Agentic co-design of low-cost robots

When does rotor cant pay on a small quadrotor co-designed with its motors and controller?

research/codesign-cant

Research programme: fixed rotor cant on a co-designed quadrotor; cant cuts attitude error by 4.2–5.3° under INDI, geometric and PID control (10 of 10 seeds) and still wins every seed in RotorPy with canted rotors; in position hold it cuts yaw error 5–8°; on a rigid frame its side force hurts one horizontal axis (1.9–3.4× the other) whichever side the centre of pressure is on; a feed-forward from the commanded torque evens the axes but adds a rare divergence…

62 tests

What it found

An open research programme (target IEEE RA-L) on when a fixed rotor cant angle pays on a small quadrotor once arm, propeller, motor headroom and attitude controller are designed with it.

Milestone 1 found that the torque limits of seven earlier simulated steps understated yaw authority by 2× to 2.8×; cant still beat no cant by 4.8° RMS.

Milestone 2 makes roll, pitch and yaw share one thrust polytope through an attitude-priority mixer (as PX4 and Betaflight do) and puts a price on hover power. Without cant the tuned vehicle saturates roll and pitch on 97% of steps and yaw gets nothing; cant's paired advantage grows to 6.08° (SD 0.28, 5 of 5 seeds). Pricing power at 0.3° per 100 W cuts hover power 4–7× for under 0.5° of error, but the chosen cant stays at 28–37° in 9 of 10 seed-arms (best of three CMA-ES starts), refuting the hypothesis of a small 5–20° optimum; cant still wins the priced objective in 5 of 5 seeds at 0.3 and 4 of 5 at 1° per 100 W. Single-start CMA-ES is unreliable on the polytope (2 of 5 seeds fell to low cant).

Milestone 4 asks whether cant only patches a weak controller: with incremental nonlinear dynamic inversion (INDI) or geometric SO(3) attitude control co-designed in place of PID (10 seeds, three CMA-ES starts each), cant still wins by 4.18° (INDI), 4.88° (geometric) and 5.27° (PID) in 10 of 10 seeds, and INDI without cant is 4.02° worse than PID with cant in 10 of 10. INDI narrows the gap by 1.1° and mostly steadies the uncanted vehicle; the chosen cant stays at a 34–35° median under every law. Simulated; TWR fixed at 2; the power price is a swept modelling choice; mass, cost and motor fits and plant ranges are inherited assumptions; no hardware.

Show the rest of the summary (37 more paragraphs, 3,285 words)

Milestone 6 adds position hold with the side force derived from the allocation (sinα / (L cosα − κ sinα) per unit roll torque, √2 smaller than an earlier study used), 6 s episodes and a 3 m/s wind step. Cant cuts yaw error by 4.7° in 6 of 6 seeds and does not buy position. Whether the side force costs position depends on the sign of the gust's disturbance torque, that is, on where the centre of pressure sits.

Under the inherited sign the force opposes the drag and costs nothing (0.96×); flipped, it costs 1.21×, past the hypothesis's 20% line (re-scored, not retrained). Cant still wins the composite in 4 of 6 and 6 of 6 seeds.

Milestone 6b derives the rigid-frame case: drag at a centre of pressure above or below the centre of mass gives opposite roll and pitch torque signs, which the inherited model did not have. Retrained under either geometry (6 seeds), the side force opposes the drag on one horizontal axis and adds to it on the other: late-window position error on the hurt axis is 1.9–3.4× the other axis. The typical design loses only 3% overall (median), but one seed in six diverges on 1–2 of 64 episodes. Cant still cuts yaw error by 5.1–7.7° (6 of 6) and wins the composite in 5 and 6 of 6 seeds. The translation model and the centre-of-pressure lever are inherited and assumed (lever swept 0–2×).

Milestone 5 re-scores the 60 milestone-4 designs, not retuned, in RotorPy with a canted-rotor wrench, motor-speed dynamics and rotor aerodynamics scaled from its Hummingbird model. Cant still wins in 10 of 10 seeds for every law under RotorPy aerodynamics (3.1–5.8°), and the six law × cant arms keep their order (ρ 0.94–1.00; ρ 0.88–0.91 over all 60 designs against 0.95–0.97 for the stylised simulator against itself). The pre-registered top-ten test fails (ρ 0.20–0.30), but the top ten are tied and the stylised simulator scores only 0.26 against itself on fresh episodes. The stylised model is 1–2° optimistic, most for INDI.

Milestone 7 lets the position loop predict the side force from its own commanded torque and lean against it (no new sensor): retrained with it, the one-axis penalty disappears (hurt/helped 1.08–1.10 against 1.91–3.41), but the term closes a loop through the motor lag that diverges on a few held-out episodes when its filter is fast (16% of episodes at 0.2 s on unretrained designs, 1.3% at 1 s), so the composite is unchanged above and worse below (J −0.33). Cant still beats no cant by 5–7° of yaw in 6 of 6 seeds.

Milestone 7b feeds forward only the attitude integrator’s torque, so the proportional and derivative path no longer closes a fast loop. Retrained with it on a frame whose centre of pressure is above the centre of mass, no held-out episode diverges at the design wind (hurt/helped 0.83; 2.06 with it switched off) and fewer do out of distribution (15 against 23 of 384); below, the designs barely use their integrator (ki/kp 0.05–0.28 s−1 against 3.7–12.6) and the term is nearly inert (ratio 2.04 against 3.01) but harmless. Added to milestone 6b’s integrator-heavy designs without retraining it diverges on 23% of episodes even with no gust torque: it must come with a retune. Cant still cuts yaw error 5–8° in 6 of 6 seeds.

Milestone 7c linearises the full attitude-plus-position loop (11 states, motor lag, one-step-old measurements, either feed-forward; matched to the simulator to 10−3) and tests it on 84 design-conditions × 64 held-out flights. It is a necessary test: no flight it calls stable at its own plant diverged (0 of 2,688 with the integrator feed-forward or none; 1 of 758 with the commanded-torque one), but 300 flights it calls unstable held on (accuracy 0.90 against 0.95 predicted).

The retrained designs’ ki/kp (3.7–4.4 s−1) sits at 0.88–1.11× the linear ceiling; the bolted-on ones at 2.3–3.4×. New designs at 0.8× their ceiling never diverge; at 1.25× the slow instability (0.5–0.7 s−1) stays bounded. Linearised about hover; the clips where divergence happens are outside the model.

Milestone 7d asks what contains those unstable flights: not the attitude integrator’s clip (never reached; removing it changes no count) and not the 6 s flight length (the linear model kills 60% of them within 6 s from their initial offset alone; none diverges in 18 s flights). The 20° tilt-command limit does: raised to 45–59° (post hoc), 83–88% of the 1.25× designs’ 18 s flights diverge, close to the 96% the model calls unstable, and designs tuned onto the ceiling lose 3–4%; at 0.8× none diverges at any limit.

Milestone 7e trains with the margin as a CMA-ES constraint (ki ≤ 0.8× ceiling at every plant corner): the centre-of-pressure-above designs land on it and lose 0.011 in held-out cost (about 0.7%); at a 45–59° tilt limit they diverge on 0 of 384 flights against 13 for the unconstrained designs, and at twice the trained lever on 4 against 15. With the centre of pressure below the constraint barely binds and changes nothing.

Milestone 7f sweeps that bound (1.0×, 0.9×, 0.8×, 0.7×, 0.6× of the ceiling; 48 new CMA-ES runs): at a 45° tilt limit the designs diverge on 9, 3, 0, 0 and 0 of 384 held-out flights (unconstrained 13), and at twice the trained lever the counts reach a floor at 0.8×; the cost penalty is about 0.011 at every bound from 0.9× down, so 0.8× is the knee and tighter margins buy nothing.

Milestone 7g asks why designs inside that margin still crash at twice the trained lever (12 and 37 of 1,152 flights at 20° and 45° limits; 45 and 55 of 384 with the centre of pressure below). Every one tips past 60° while the attitude-priority mixer is saturated and the gust torque nears the motors' limit. An idealised mixer removes two thirds; with 1.25× torque room (1.5× below), all. A wider crash bound, halved position gains or an airmode mixer do not, so the residual is a thrust-headroom limit (thrust-to-weight 2 is assumed), not a stability one.

Milestone 7h makes thrust-to-weight a design variable. On the co-designed airframes the 15-minute endurance sizing closes only up to 2.1–2.5: re-sized there, 7e’s designs lose every doubled-lever crash (4 and 12 → 0 above; 45 and 55 → 0 and 2 below) but weigh 78% more and hover on 2.4× the power. Retrained with thrust-to-weight free (1.5–4), all 12 seeds pick 2.05–2.55 and spend the headroom on larger propellers and lower hover power, not torque, so the doubled-lever crashes stay (9 against 4 above, 38 against 45 below).

Milestone 7i trains across centre-of-pressure levers 0–2. Where the old designs were fragile (below), doubled-lever crashes fall from 45 to 14–20 of 384 and tripled-lever crashes from 159 to 86–87, at no cost at the design lever; where they were robust (above), nothing changes. Free thrust-to-weight adds nothing: what the tuner buys is mass, which at fixed thrust-to-weight brings bigger motors while the gust torque stays put (post hoc, mass ranks with tripled-lever crashes at ρ = −0.87 above and −0.94 below; thrust-to-weight at +0.08 and −0.02).

Milestone 7j explains the crashes the retrained designs show with no gust torque at PX4’s 45° tilt limit (24 of 1,152 flights, 0 for 7e): every one is a tip-over in the first 0.8 s, recovering from the 0.5 m start offset with a position gain at its search bound; the 20° limit used in training clipped that lean. Halving the offset, an ideal mixer or halved position gains remove all of them. Tune at the tilt limit you fly.

Milestone 7k fixes it in the controller instead: capping the position error in the proportional term at 0.25–0.35 m (the 20° lean the tune saw) or ramping the setpoint removes every transient crash at 45° (34 → 0 of 3,072 flights; 935 → 0 with a 1.0 m offset) for 1–2% of held-out cost at 20°; a tilt-rate limit of 0.5–1 rad/s makes it far worse (up to 515), and PX4’s tilt-magnitude cone changes nothing at the trained offset.

Milestone 7l retrains at 45°: alone it lands on the 20° tune’s cost and its crashes at a 1.0 m offset (92–103 of 384, which the cap removes); with the cap and the position-gain box widened from 10 to 31.6 s−2, the tuner picks gains of 22–31 s−2 and a 11–19% lighter airframe with 18–30% less hover power at the same held-out tracking, but those designs crash on 267–274 of 384 flights without the cap and on 73–105 (against 12–52) at twice the centre-of-pressure lever.

Milestone 7m adds the lever-aware objective to that wide box: the tuner still picks 20–31 s−2, but its designs weigh 1.40–1.45 kg (between the lever-1 wide designs’ 1.10–1.29 and the old box’s 1.53–1.60) and have the longest frame arms and highest torque per weight of any set, track best at the trained levers, and crash on 52–53 of 384 flights at twice the lever (against 73–105), but 231–250 without the cap.

Milestone 7n pins the arm at the lever-1 designs’ length and retrains: the lever-aware designs crash on 39–46 of 384 at twice the lever, no more than with the arm free, reaching it through mass and gains; on a fixed design, a 4–7 mm longer arm cuts those crashes by about a fifth (73→59, 105→87) for 4% more mass, since frame mass grows with the arm.

Milestone 7o sweeps the arm from 85 to 135 mm on 48 fixed designs: crashes at twice the lever fall monotonically with the arm wherever the design stays valid (lever-1 designs 73→6 and 105→44 of 384), with no interior optimum; the limits are the stability margin below the tuned length and the mass–battery spiral above it (+26% to +98% mass by 120–135 mm); the growing gust exposure takes back up to a third of the gain; and on the lever-1 tune a longer arm alone matches the lever-aware designs’ robustness at the same mass, power and tracking. The motor-mass and sizing fits are inherited assumptions.

Milestone 7p sources the frame-mass law from 36 listed commercial frames (one retailer, verbatim spec lines): stretching one frame design adds 0.51 kg per metre of arm (0.26 g per mm of wheelbase), 64% of the assumed 0.8, while across brands a frame designed for a larger size weighs 2.27 kg/m more. Re-swept under these laws without retraining, crashes still fall monotonically with the arm, but with the stretched-design slope much more slowly (lever-1 designs, centre of pressure below: 66 against 44 of 384 at 135 mm), and the arm stays valid to 150 mm instead of 130 (above).

At a fixed arm the heavier law crashes less on 253 of 350 design-points, so part of 7o’s arm effect was mass. The inherited endurance closure multiplies frame mass 6–15×: the cross-brand law’s 20–43 g lighter frames make the sized vehicles 0.18–0.64 kg lighter, and on the same tunes they crash on 151–160 of 384 flights instead of 73–105.

Milestone 7q checks the endurance sizing closure itself. Solved exactly instead of by a damped iteration stopped after 60 steps, 28 of the 48 tuned designs, including all six heavy (≥ 1.7 kg), near-crash-free ones, have no fixed point at 15 minutes: the tuners sat every design on the closure’s edge. Listed components move all three inherited constants: LiPo packs store 132 Wh/kg (assumed 150; 140 packs), 5-inch propulsion converts 0.426 of electrical power into ideal momentum power (assumed 0.4875; SunnySky bench tables), and motors weigh 0.58 of the assumed law (33 motors).

With all three, no design closes on LiPo (longest closing endurance 11.9–13.0 min); on Li-ion (210 Wh/kg) all close at about 0.75 kg, and on their old tunes they crash on 20–33 more of 64 flights per design at twice the centre-of-pressure lever. Listings, not measurements. Retuned on the sourced Li-ion closure (milestone 7r, 48 CMA-ES runs), the tuners do not keep the 0.75 kg vehicle: they shrink the propeller (0.13 → 0.11 m), draw more hover power and size back to 0.99–1.29 kg on the closure’s edge, and robustness again follows mass on one curve: the lever-aware objective crashes on 54 and 64 of 384 flights at twice the lever and a 45° tilt limit against 106 and 131 for the lever-1 objective, and is 14–24% heavier.

The mass-robustness relation is not an artefact of the truncated solve; the inherited power price and fixed 15-minute endurance now set the mass. Swept over nine motor-lag laws (7.5–30 ms at 5 inches, diameter exponent 0–3; milestone 7s, not retuned), the orderings survive (lever-aware beats lever-1 under all nine; ρ(mass, crashes) −0.68 to −0.93 up to 15 ms), but the counts do not: with one 15 ms lag for every propeller the retuned lever-aware designs crash on 85 and 92 of 384 flights against 51 and 53 for the published ones, only 5 of 24 retuned designs pass the stability-margin test, and at 30 ms one design of 72 does.

The lag law is assumed, not sourced. Taken from two published system identifications of 5–5.1 inch quadrotors (33 and 39.1 ms; milestone 7t), the lag is 2.4× the assumed value: at it none of the 24 designs tuned at 15 ms passes the margin test and they crash on 359–384 of 384 flights, while designs retuned at 36 ms all pass, crash on 63–161, keep the lever-aware advantage (+14–28% mass), and crash less than the 15 ms tunes even at 15 ms (8–59 against 54–131). The two identifications come from one lab and the diameter exponent is not sourced.

Milestone 7u retunes at the slower identification (39.1 ms) instead of the midpoint: inside the identified range those designs crash less than the 36 ms ones under all three lags in all four arm-geometries, even at 36 ms (109, 132, 61, 59 against 149, 161, 63, 82 of 384), all 24 stay certified at 33–39.1 ms (20 of 24 midpoint designs do), at the same mass. Every cell leans the same way but none is significant at six seeds (Mann–Whitney p 0.08–0.81); the price is a lost certificate on 4 of 6 designs of one set at the old 15 ms law. With only a range for the motor lag, tune at its slow end.

Milestone 7v randomised the training lag across 33–39.1 ms instead (certified at both ends): those designs crash more than the slow-end ones in 11 of 12 law × arm-geometry cells, including all four at 33 ms (101, 112, 61, 53 against 80, 107, 37, 46 of 384), and land between the midpoint and slow-end tunes; not significant per cell at six seeds (sign test p = 0.125). Train on plants no faster than the slow end.

Milestone 7w tuned slower still, at 45 ms (unsourced, beyond the slower identification): those designs crash less than the 39.1 ms tunes in 19 of 20 law × arm-geometry cells (one tie), including all four at 33 ms (75, 77, 27, 34 against 80, 107, 37, 46 of 384), 21–29% fewer over the identified range, at the same mass; 22 of 24 stay certified across 33–39.1 ms (the two failures sit on the boundary, ρ ≤ 1.0002). Sign test p = 0.125 at every lag. No knee yet: tune slower than the slowest measurement, limited by the certificate at the fast end.

Milestone 7x tuned the lever-aware objective at 52 and 60 ms (unsourced floors), with the certificate also required at 33 ms: crashes over the identified range keep falling (348, 236, 162, 87 of 2304 for 39.1, 45, 52 and 60 ms tunes; 8 and 12 against 27 and 34 of 384 at 33 ms for 60 against 45 ms), the 60 against 45 ms gap is measured (Mann–Whitney p = 0.004–0.014 in all six identified cells), and all 24 designs stay certified, the 33 ms corner slack on 22 of them.

No knee by 60 ms; the price is 4–8% more mass, a 20% softer position loop and about 5% worse tracking, so the floor is a mass and tracking choice, not a stability one. Simulated; the stress test (twice the trained lever, 45°) favours a slow, heavy vehicle.

Milestone 7y prices that softer loop without retraining: flown through a 1 m position step and a calm-to-wind step at the design lever, the 39.1, 45, 52 and 60 ms tunes rise to 90% in 1.44, 1.58, 1.64, 1.74 s (centre of pressure above; 1.39–1.77 s below) and overshoot 6.4, 5.9, 3.8, 3.2 cm (9.7–2.7 below), monotone in every cell; settle time to 5 cm barely moves (+16% above, −24% below for 60 against 45 ms), and peak drift after a 6 m/s wind step grows by 4–5 mm. In calm air the motor lag itself changes nothing; 8 of 48 designs carry a slow final approach (position gain 4.5–9.7 s−2) whatever the floor.

Milestone 7z adds a 75 ms floor: inside the identified lags it crashes on 29, 33 of 1152 stress flights against the 60 ms tune's 36, 51 (not separated at six seeds in five of six cells), and clearly fewer at slower flown lags (49, 45 against 105, 83 at 60 ms; p = 0.010–0.026 in three of four cells at 60 and 75 ms); the designs are lighter (1.43 against 1.47–1.51 kg) and all certified, but rise 5–14% slower, overshoot more (4.8, 3.2 against 3.2, 2.7 cm) and settle later, so 60 ms stays the tune for motors near the identifications.

Milestone 7aa separates the tuning floor from the ceiling: trained at the measured 39.1 ms floor with lags up to 150 ms, the designs crash on 27, 38 of 1152 identified-lag stress flights against 170, 178 for the 39.1–78.2 ms tune (p = 0.002–0.005 in all 12 cells) and are not separated from the 75 ms tune anywhere, with no faster rise (2.23, 1.76 s): the robustness of the slower tunes came from covering slow plants, not from leaving out fast ones.

Milestone 7ab sweeps that ceiling at the same floor: 100 ms leaves 82, 109 identified-lag crashes (more than 150 ms in 11 of 12 cells, p = 0.002–0.046), while 200 ms is not separated from 150 ms inside the measured lags (17, 36) and crashes far less at a 75 ms motor (50, 71 of 384 against 138, 131); the price is rise time (1.47, 1.54 s at 100 ms; 2.02, 2.27 s at 200 ms), 5–7% more mass and a slower final approach, so the ceiling is a dial between robustness to unmeasured slow motors and response, with the knee between 100 and 150 ms (about 4× the measured lag).

Milestone 7ac places that knee with a 125 ms ceiling: identified-lag crashes 39, 62 (between 100 ms’s 82, 109 and 150 ms’s 27, 38, separated from 150 ms in 1 of 6 cells), rise 1.57, 1.65 s (faster than 150 ms above, p = 0.026), but at a 75 ms motor 204, 206 of 384 crashes against 138, 131 (p = 0.010, 0.011): the curve is a slope, not a cliff, slow-motor robustness falls almost linearly with the ceiling, and 5 of 12 125 ms designs never settle within 5 cm. Simulated; the ceilings are brackets, not sourced.

Milestone 7ad traced the designs that never settle on a 1 m step (18 of 72) to the position integrator’s slow pole near −ki/kp: a three-state outer-loop model calls all 96 untested designs correctly (r 0.9995), and integrating only within 0.25 m of the set point removes every unsettled design and halves overshoot with stress crashes unchanged (819 against 820), for a 10% slower rise (simulated).

Research log

Ongoing: each run adds a batch and an entry, newest first. Source: LOG.md.

2026-11-19 · Milestone 7ad: why do a quarter of the tunes never settle on a 1 m step? The integrator's slow pole, and conditional integration as a fix without retuning (predictions written and committed before experiments/m7ad_settle/run.py was run at full budget; outcome below)

### Outcome (experiments/m7ad_settle/verdicts.txt, results.json; part A 288 step flights, 84 s; part B 1,944 response flights, 545 s, and 1,152 stress evaluations, 1,283 s, on 4 cores; figure m7ad_settle.png; simulated)

Show 37 earlier entries

2026-11-19 · Paper step: section VII of the draft, how wide the motor-lag randomisation should be (milestones 7s-7ac), with pooled tests

  • Why: pick_track.py flagged 22 log entries since a paper step and 29 sub-steps of milestone 7, each turning one knob; the draft still ended at milestone 7g. Eleven training ranges (7t-7ac) had each been read with six seeds per geometry and per-cell tests that rarely separated neighbours, so the section needed one table and tests with more power, not another knob.
  • Built: paper/DRAFT.md section VII written in full (setting from 7s-7t, Table VII-1 of all eleven training ranges, five findings, a builder rule, a figure plan); abstract, contributions, related work, limitations, conclusion and to-do updated; old VII-VIII renumbered VIII-IX. paper/check_numbers.py now generates every Table VII-1 row from the results files (--table7 prints them) and checks 13 more quoted figures (58 checks). paper/table7_stats.py (stdlib, post hoc, runs no flights) writes paper/table7_stats.txt: per-design identified totals, stratified bootstrap 95% intervals, a Jonckheere-Terpstra trend test along the ceiling sweep and pooled (12 designs a side) permutation Mann-Whitney tests. Five references added (two motor identifications already in data/motor_lag/, three on domain randomisation; arXiv ids checked by cite_check.py). Two tests: the stats file is reproduced exactly, and a row sums its three laws.
  • Not predicted (post hoc; read as such): pooled over geometries, the 33-78.2 ms draw against the 39.1-78.2 ms one, 486 against 348 of 2,304, p = 0.011 (per cell it never reached p < 0.05); the 39.1 ms floor against the 75 ms floor at a 150 ms ceiling, 65 against 62, p = 0.83; along the ceilings 78.2-200 ms a decreasing trend at the permutation floor (p = 0.00005) in both geometries and at the 75 ms law; neighbours separated at 78.2→100 (p = 0.018) and 100→125 ms (p = 0.025), not at 125→150 (0.10) or 150→200 ms (0.16); at 75 ms every neighbour pair separated (p ≤ 0.0002).
  • Reading: pooling turns 7v's "probable" loss from drawing fast plants into a measured one, and 7aa's "floor does not matter at a high ceiling" into a measured null; the knee in identified-range robustness is at about 125 ms (3.2× the slow identification), as 7ac's slope reading said. The section's mechanism paragraph (the mean over episodes trades slow-edge margin for fast-plant tracking) is labelled a reading, not tested.
  • Limits: pooling the geometries treats them as exchangeable design sets; the tests are post hoc and many; six seeds per geometry; ceilings are brackets, not sourced.
  • Next: figures VII-a and VII-b; the same coverage sweep on torque scale and mass; a settle-time term in the objective.

2026-11-19 · Milestone 7ac: where is the knee? A 125 ms training ceiling at the identified floor (predictions written and committed in 118e650fc before experiments/m7ac_ceil125/run.py was run at full budget; outcome below)

  • Why: milestone 7ab's second next step. 7ab put the knee in identified-range robustness between 100 and 150 ms (identified total 82, 109 at 100 ms against 27, 38 at 150 ms) and the rise-time jump in the same interval (1.47, 1.54 s against 2.23, 1.76 s on a 1 m step at the 36 ms law). Two questions a builder needs answered: is the crash drop a cliff or a slope across that interval, and does the rise-time price arrive with the robustness or before it? If 125 ms keeps most of 150 ms's robustness at 100 ms's rise time, the recommendation becomes "about 3× the measured lag"; if not, 4× stands.
  • Plan (experiments/m7ac_ceil125/run.py): 7ab's training exactly, through 7ab's own patch and train (closure src_li, 45° tilt limit, A_P cap, wide kp box, integrator feed-forward, margin constraint, TWR 2, 3200 CMA-ES evaluations, 6 seeds × 2 starts × 2 geometries) at the 39.1 ms D<sup>0</sup> floor with plant lag × U(1, 125/39.1) (lag39c125); certificate corners 33, 39.1 and 125 ms. 125 ms is not sourced: it is the midpoint of 7ab's bracket. Same random stream, so tau and mass draws are 7ab's. Scored on 7z's held-out stress set (seed 9999, 64 episodes, four capped conditions, margin test, eight laws) and 7y's response tests (1 m step, 3 and 6 m/s wind steps; lever 1, 45°, calm, seed 9999) for six sets: lag39 (c = 78.2), lag39c100, lag39c125, lag39c150, lag39c200, lag75. All old designs are re-scored, which is the reproduction check.
  • Looked at before writing these: 7ab's results; a smoke run (64 evaluations, 2 seeds, 8 episodes on seed 4242, never seed 9999) to check that the patch gives corners 33.0, 39.1, 125.0 ms and training lags 39.1-125.0 ms and that the script runs end to end. Its designs are unconverged and say nothing about the questions.
  • Predictions (identified total = L2-45 diverged flights at lever 2, 45°, cap, summed over 6 seeds and over the 33, 36 and 39.1 ms laws, of 1152; values as above, below; lag39c100 82, 109; lag39c150 27, 38): - AC1 (prior 55%): a slope, not a late cliff: lag39c125's identified total is at most the midpoint of lag39c100 and lag39c150 (54.5, 73.5) in both geometries. - AC2 (prior 30%): 125 ms is enough: lag39c125's identified total is within 10 of lag39c150's (at most 37, 48) in both geometries. - AC3 (prior 55%): the response price comes with the coverage: median rise time to 90% of the 1 m step at the 36 ms law, lag39c125 lies strictly between lag39c100 (1.471, 1.536 s) and lag39c150 (2.231, 1.757 s) in both geometries. - AC4 (prior 70%): slow-motor robustness is monotone in the ceiling: at the 75 ms law lag39c125's L2-45 count lies strictly between lag39c150's (138, 131) and lag39c100's (266, 277) in both geometries. - AC5 (prior 75%): the certificate holds: at least 11 of 12 lag39c125 designs pass the margin test under all three identified laws. - AC6 (prior 50%): mass is monotone too: mean mass of lag39c125 lies between lag39c100's (1.423, 1.331 kg) and lag39c150's (1.451, 1.422 kg) in both geometries.
  • Reading fixed in advance: AC1 and AC2 held with AC3 refuted on the fast side (rise at or below the midpoint of 100 and 150 ms's) would move the builder recommendation to about 3× the measured lag. AC1 held with AC2 refuted says the curve is a slope and the choice inside 100-150 ms is a smooth trade, so the recommendation stays "at least 4×" for robustness first. AC1 refuted says the drop happens late, close to 150 ms. AC4 refuted says the slow tail is not a simple function of the ceiling.
  • Analysis: counts per set, geometry and law; margin validity; rejected share; mass, propeller, kp; response per set and law. Post hoc only: Mann-Whitney U per cell on six per-seed counts and per-design response medians, lag39c125 against lag39c100 and against lag39c150. Checks: the five old sets reproduce 7ab's counts and responses. Six seeds, one evaluation seed.

### Outcome (experiments/m7ac_ceil125/results.txt, verdicts.txt, posthoc.txt; 24 CMA-ES runs of 3200 evaluations, 4358 s on 4 cores, then 72 designs × 8 laws × 4 conditions × 64 episodes, 685 s, and the response tests on all 72; figure m7ac_ceil125.png; simulated)

  • A slope, not a cliff. Identified total (of 1152, above, below) by ceiling 78.2 / 100 / 125 / 150 / 200 ms: 170 / 82 / 39 / 27 / 17 and 178 / 109 / 62 / 38 / 36. AC1 held (39, 62 against midpoints 54.5, 73.5); AC2 refuted (39, 62 against 37, 48: narrowly above, clearly below). Post hoc, lag39c125 is not separated from lag39c150 in 5 of 6 identified-range cells (p = 0.13-0.53; 39.1 ms below p = 0.040), nor from lag39c100 in any (p = 0.075-0.25), with six seeds.
  • Slow-motor robustness is close to linear in the ceiling. L2-45 at the 75 ms law (of 384): 331 / 266 / 204 / 138 / 50 above and 339 / 277 / 206 / 131 / 71 below. AC4 held: lag39c125 sits between its neighbours and is separated from both (p = 0.010-0.015). At 60 ms 92, 99 against 56, 55 (p = 0.064, 0.063) and 170, 197 (p = 0.053, 0.013).
  • Response: most of 100 ms's speed. Rise to 90% of a 1 m step at the 36 ms law 1.57, 1.65 s against 1.47, 1.54 (100 ms) and 2.23, 1.76 (150 ms). AC3 held; faster than 150 ms above (p = 0.026), not separated below or from 100 ms. Overshoot 8.6, 5.9 cm (100 ms 9.4, 5.2; 150 ms 3.6, 3.5). But 5 of 12 lag39c125 designs never settle within 5 cm on most step flights (last-second distance 6-8 cm, response position kp 5.0-6.6), the most of any set (100 ms 2, 150 ms 4, 200 ms 3); the above set's censored median settle time is infinite.
  • Certificate, mass. 12 of 12 pass the margin test under the three identified laws (AC5); rejected share 0.13, 0.13. Mass 1.492, 1.402 kg, above 150 ms's 1.451 above and between 100 and 150 ms below: AC6 refuted; mass is not monotone in the ceiling at six seeds (per-seed spreads are about 0.1 kg).
  • Verdicts. AC1 held. AC2 refuted. AC3 held. AC4 held. AC5 held. AC6 refuted. Checks: all five old sets reproduce 7ab's stress counts and step-response rise times exactly.
  • Reading (against the readings fixed in advance): AC1 held with AC2 refuted is the "slope" reading: there is no sharp knee between 100 and 150 ms; identified-range crashes fall steeply to about 125 ms and slowly after, while slow-motor crashes fall about linearly all the way to 200 ms. AC3 held, with the rise near 100 ms's, means most of the response price arrives between 125 and 150 ms above (1.57 to 2.23 s), but below it grows smoothly (1.54, 1.65, 1.76 s). For a builder: if the motors are known to sit near the measured lag, about 3× (125 ms) gives nearly 150 ms's robustness at the measured lags with 100 ms's rise time, at the cost of a slow final approach in some seeds; if the motors may be slower than measured, every extra millisecond of ceiling buys slow-motor robustness at a steady rate, and 4× or more stands.
  • Limits: one arm; ceilings are brackets, not sourced; six seeds and one evaluation seed cannot separate neighbouring ceilings inside the identified range; response tests are calm air at the design lever; the settle-time result counts designs, not a predicted quantity.
  • Next: a settle-time or position-kp floor as a training constraint (7ab's first next step; 7ac makes it sharper, since the 125 ms tune's unsettled designs are what keeps it from being the default); twelve seeds at 125 and 150 ms to separate them inside the identified range; the same coverage test on the torque-scale and mass ranges.

2026-11-18 · Milestone 7ab: how much lag coverage is enough? The training-ceiling sweep at the identified floor, c = 100 and 200 ms (predictions written and committed in 4cd7341ce before experiments/m7ab_ceilsweep/run.py was run at full budget; outcome below)

  • Why: milestone 7aa's first next step. 7aa found the training ceiling, not the floor, buys the slow tunes' robustness: at the identified 39.1 ms floor a 150 ms ceiling cut identified-range crashes from 170, 178 (7u, ceiling 78.2 ms) to 27, 38, level with the 75 ms tune (29, 33). Two points (78.2 and 150 ms) do not say where coverage stops paying, nor whether more coverage starts to cost, so a builder still cannot say how wide the range should be.
  • Plan (experiments/m7ab_ceilsweep/run.py): 7aa's arm and training exactly (closure src_li, 45° tilt limit, A_P cap, wide kp box, integrator feed-forward, margin constraint, TWR 2, 3200 CMA-ES evaluations, 6 seeds × 2 starts × 2 geometries) at the 39.1 ms D<sup>0</sup> floor with plant lag × U(1, c/39.1) for c = 100 ms (lag39c100) and c = 200 ms (lag39c200); certificate corners 33, 39.1 and c ms. Neither ceiling is sourced: they bracket 7aa's 150 ms on each side. Same random stream, so tau and mass draws are 7aa's. Scored on 7z's held-out stress set (seed 9999, 64 episodes, four capped conditions, margin test, eight laws) and 7y's response tests (1 m step, 3 and 6 m/s wind steps; lever 1, 45°, calm, seed 9999) for all five sets: 7u's lag39 (c = 78.2), lag39c100, 7aa's lag39c150, lag39c200, 7z's lag75.
  • Looked at before writing these: 7aa's results; a smoke run (64 evaluations, 2 seeds, 8 episodes on seed 4242, never seed 9999) to check the patch and that the script runs end to end. Its designs are unconverged and say nothing about the questions.
  • Predictions (identified total = L2-45 diverged flights at lever 2, 45°, cap, summed over 6 seeds and over the 33, 36 and 39.1 ms laws, of 1152; values as above, below; lag39 170, 178; lag39c150 27, 38; lag75 29, 33): - AB1 (prior 60%): 100 ms buys most of it: lag39c100's identified total is at most the midpoint of lag39 and lag39c150 (98.5, 108) in both geometries. - AB2 (prior 35%): 100 ms is enough: lag39c100's identified total is within 10 of lag39c150's (at most 37, 48) in both geometries. - AB3 (prior 45%): more coverage keeps paying at slow motors: at the 75 ms law lag39c200 crashes on fewer L2-45 flights than lag39c150 (138, 131) in both geometries. - AB4 (prior 45%): more coverage costs inside the identified range: lag39c200's identified total exceeds lag39c150's (27, 38) in both geometries. - AB5 (prior 60%): coverage costs response: median rise time to 90% of the 1 m step at the 36 ms law, lag39c100 < lag39c200 in both geometries. - AB6 (prior 70%): the certificate holds: at least 22 of the 24 new designs pass the margin test under all three identified laws. - AB7 (prior 55%): the slow tail grows with the ceiling: more lag39c200 designs than lag39c100 designs have position kp below 8 (7aa's slow-tail designs had 5.0 and 6.2), summed over geometries.
  • Reading fixed in advance: AB1 and AB2 held say a modest ceiling (about 2.5× the identified lag) is enough, and the recommendation becomes "train to 100 ms"; AB1 held, AB2 refuted puts the knee between 100 and 150 ms. AB3 held with AB4 refuted says coverage is close to free up to 200 ms, so the builder should err wide; AB4 held says there is an optimum ceiling and 7aa's 150 ms sits near it. AB5 and AB7 say what the coverage costs in the builder's units (rise time, slow designs).
  • Analysis: counts per set, geometry and law; margin validity; rejected share; mass, propeller, kp; response per set and law. Post hoc only: Mann-Whitney U per cell on six per-seed counts against lag39c150. Checks: the three old sets reproduce 7aa's counts and responses. Six seeds, one evaluation seed.

### Outcome (experiments/m7ab_ceilsweep/results.txt, verdicts.txt, posthoc.txt; 48 CMA-ES runs of 3200 evaluations, 7167 s on 4 cores, then 60 designs × 8 laws × 4 conditions × 64 episodes, 485 s, and the response tests; figure m7ab_ceilsweep.png; simulated)

  • 100 ms is not enough; coverage pays in steps up to 150 ms, then flattens inside the identified range. Identified total (of 1152, above, below) by ceiling 78.2 / 100 / 150 / 200 ms: 170 / 82 / 27 / 17 and 178 / 109 / 38 / 36 (lag75 29, 33). Post hoc, lag39c100 crashes more than lag39c150 in 11 of 12 cells from 33 to 75 ms (p = 0.002-0.046; 33 ms above p = 0.074). lag39c200 is not separated from lag39c150 in any identified-range cell (p = 0.07-1.0). AB1 refuted by one crash (below 109 against a midpoint of 108; above 82 against 98.5); AB2 refuted.
  • More coverage keeps paying at slow motors, and does not cost inside the range. At the 75 ms law lag39c200 crashes on 50, 71 of 384 against lag39c150's 138, 131 (p = 0.005, 0.020) and lag75's 117, 103; at 60 ms 26, 41 against 56, 55 (p = 0.004, 0.41). AB3 held, AB4 refuted (17, 36 against 27, 38).
  • The price is response and mass. Rise to 90% of a 1 m step at the 36 ms law: 1.47, 1.54 s at 100 ms (as fast as lag39's 1.44, 1.39; faster than lag39c150, p = 0.004, 0.037), 2.23, 1.76 s at 150 ms, 2.02, 2.27 s at 200 ms (below slower than lag39c150, p = 0.041). AB5 held. The 200 ms designs are heavier (1.546, 1.525 kg against 1.451, 1.422) with lower position kp (response-test median 4.5, 5.5 against 6.3, 5.8 at 150 ms and 7.5, 6.7 at 100 ms), and 3 of 6 above designs never settle within 5 cm (none dead); peak drift after a 6 m/s wind step above 4.3 cm against 3.2 (p = 0.015).
  • Certificate and training. All 24 new designs pass the margin test under the three identified laws (AB6). Rejected share 0.13, 0.10 (100 ms) and 0.20, 0.17 (200 ms) against 0.15, 0.13 at 150 ms: the widest interval is the hardest to certify, unlike 7aa's 150 ms.
  • AB7 and a scoring correction. The committed verdicts.py read the stress run's pos_kp field, a different quantity from the response test's position kp from which the prediction's reference values (7aa's slow-tail designs at 5.0 and 6.2) came; on that field no design in any set is below 8 and AB7 printed refuted (0 against 0). Corrected to the response field after the run, before any reading was written: AB7 held, 12 against 9. The threshold was poorly chosen: 8 to 12 designs of every set are below 8, including 8 of lag39's. The median kp (above) falls with the ceiling, which is the trend AB7 meant; designs never settling within 5 cm on most step flights are 2 (100 ms), 4 (150 ms), 3 (200 ms), no trend.
  • Verdicts. AB1 refuted (narrowly). AB2 refuted. AB3 held. AB4 refuted. AB5 held. AB6 held. AB7 held on the corrected field, with the threshold caveat above. Checks: lag39, lag39c150 and lag75 reproduce 7aa's stress counts and responses exactly.
  • Reading (against the readings fixed in advance): AB1 and AB2 refuted put the knee in identified-range robustness between 100 and 150 ms, about 2.5-4× the identified lag. AB3 held with AB4 refuted says coverage above that is free inside the identified range and buys a lot at slow motors (the 200 ms tune is the most robust at 60-75 ms of every tune in the programme), so there is no interior optimum by crash count. The price is paid in rise time, mass and a slower final approach, so the ceiling is a dial between robustness to unmeasured slow motors and response, not a free choice. For a builder: train at least to about 150 ms (4× the measured lag); go to 200 ms only if motors may genuinely be that slow, at about 0.3-0.5 s of rise time and 5-7% of mass.
  • Limits: one arm; three ceilings, none sourced (brackets); six seeds, one evaluation seed; response tests are calm air at the design lever; the 200 ms tune's slower approach is a training-objective trade, not a stability failure.
  • Next: a position-kp or settle-time floor as a training constraint, now that the slowest approach comes with the widest coverage (would it keep the 200 ms tune's slow-motor robustness at 150 ms's response?); a ceiling at 125 ms to place the knee; the same coverage test on the torque scale and mass ranges.

2026-11-18 · Milestone 7aa: separate the tuning floor from the ceiling; is the 75 ms tune's robustness bought by excluding fast plants or by covering slow ones? (predictions written and committed before experiments/m7aa_ceiling/run.py was run at full budget; outcome below)

  • Why: milestone 7z's first next step. 7u-7z raised the tuning floor F with the plant randomisation fixed at lag × U(1, 2), so every step moved both ends of the training lag interval [F, 2F] and its certificate corners together: 7u's 39.1 ms tune trained on 39.1-78.2 ms, 7z's 75 ms tune on 75-150 ms. Because the lever-aware arm's attitude loop is a PID with no internal lag model, the floor enters training only through that interval (and the certificate corners), so the two readings are distinct and testable: the slower tunes win either because their training covers slow plants (the ceiling) or because it leaves out the fast ones the search would otherwise favour (the floor).
  • Plan (experiments/m7aa_ceiling/run.py): 7z's arm exactly (closure src_li, 45° tilt limit, A_P cap, wide kp box, integrator feed-forward, margin constraint, TWR 2, 3200 CMA-ES evaluations, 6 seeds × 2 starts × 2 geometries) with the floor at the identified 39.1 ms D<sup>0</sup> (data/motor_lag/) and the plant lag × U(1, 150/39.1 = 3.836), so training lags span 39.1-150 ms; certificate corners 33, 39.1 and 150 ms. The 150 ms ceiling is not sourced: it is 7z's, kept so the only change against 7z is the floor. The plant draw keeps its random stream (lag is the last of the three uniforms), so tau and mass draws are 7z's numbers. Scored on 7z's held-out stress set (seed 9999, 64 episodes, four capped conditions, margin test, eight laws) with 7u's lag39, 7x's lag60 and 7z's lag75 designs alongside, and on 7y's response tests (1 m step, 3 and 6 m/s wind steps; lever 1, 45°, calm, seed 9999, 64 episodes) for lag39, lag39c150 and lag75.
  • Looked at before writing these: a smoke run (64 evaluations, 2 seeds, 8 episodes on seed 4242, not seed 9999) to check the patch (corners 33.0, 39.1, 150.0 ms; training lags 39.1-150.0 ms over 4096 draws; ρ at x = 0.5 1.0297 against 7z's 1.0302) and that the script runs end to end. Its designs are unconverged (97% of candidates rejected) and say nothing about the questions.
  • Predictions (L2-45 = diverged flights at lever 2, 45°, cap, summed over 6 seeds, of 384; identified total = L2-45 summed over 33, 36 and 39.1 ms, of 1152; values as above, below; lag39 170, 178; lag60 36, 51; lag75 29, 33): - AA1 (prior 55%): the ceiling buys most of it: lag39c150's identified total is at most the midpoint of lag39 and lag75 (99, 105) in both geometries. - AA2 (prior 30%): the ceiling buys all of it: lag39c150's identified total is at most lag60's (36, 51) in both geometries. - AA3 (prior 60%): at slow motors the floor still matters: at the 75 ms law lag39c150 crashes on more L2-45 flights than lag75 (117, 103) in both geometries. - AA4 (prior 70%): the certificate holds: at least 11 of 12 lag39c150 designs pass the margin test under all three identified laws. - AA5 (prior 55%): the wide tune keeps a faster response: median rise time to 90% of the 1 m step, lag39c150 < lag75 (1.98, 1.86 s) in both geometries. - AA6 (prior 50%): the wider interval is harder to certify in training: mean rejected share lag39c150 > lag75 (0.15, 0.16) in both geometries.
  • Reading fixed in advance: AA1 held says the slow-floor result of 7u-7z is mostly a coverage result: train across the whole plausible lag range anchored at the identified lag rather than at a slow nominal one, and with AA5 held that costs less rise time than the 75 ms tune. AA1 refuted says leaving out the fast plants is what helps (the search, scored on a mean over episodes, trades slow-plant robustness for fast-plant tracking when both are present), so the floor itself is the design choice and 7z's recommendation stands. AA3 held with AA1 held splits it: coverage buys the identified range, concentration buys the slow tail.
  • Analysis: counts per set, geometry and law; margin validity; rejected share; mass, propeller, kp; response metrics per set and law. Post hoc only: Mann-Whitney U per cell on the six per-seed counts, lag39c150 against lag39 and against lag75. Checks: the three old sets reproduce 7z's (lag60, lag75) and 7x's (lag39) counts and 7y's and 7z's responses. Six seeds, one evaluation seed.

### Outcome (experiments/m7aa_ceiling/results.txt, verdicts.txt, posthoc.txt; 24 CMA-ES runs of 3200 evaluations, 4608 s on 4 cores (the wider lag range makes each run slower than 7z's), then 48 designs × 8 laws × 4 conditions × 64 episodes, 484 s, and the response tests; figure m7aa_ceiling.png; simulated)

  • The ceiling buys essentially all of it. Identified total (L2-45 at 33 + 36 + 39.1 ms, of 1152): 27, 38 for the 39.1 ms floor with a 150 ms ceiling, against 170, 178 for the 39.1 ms floor with its 78.2 ms ceiling (7u), 36, 51 for the 60 ms tune and 29, 33 for the 75 ms tune (above, below). AA1 and AA2 held. Post hoc, against lag39 the wide tune crashes less in all 12 law × geometry cells from 33 to 75 ms (Mann-Whitney p = 0.002-0.005), about a fifth as often inside the identified range; against lag75 it is not separated in any cell (p = 0.38-1.0).
  • A small residue of the floor shows only at slow motors, and it is not separated. At the 60 and 75 ms laws the wide tune crashes on 56, 55 and 138, 131 flights against lag75's 49, 45 and 117, 103 (AA3 held on the stated count rule; post hoc p = 0.38-0.69). Inside the identified range and at 45 ms the two tie (7 against 8 ... 21 against 18).
  • Including the fast plants did not buy back the response (AA5 refuted). Rise to 90% of a 1 m step 2.23, 1.76 s against lag75's 1.98, 1.86 and lag39's 1.44, 1.39; overshoot 3.6, 3.5 cm (lag75 4.8, 3.2; lag39 6.4, 9.7). Against lag39 the slower rise is separated (p = 0.015, 0.009); against lag75 neither rise nor overshoot is (p = 0.24-0.70). With the centre of pressure below, the wide tune sits on 7z's frontier and dominates the 60 ms tune (38 crashes at 1.76 s against 51 at 1.77 s); above, it lies right of it, pulled by two slow-tail designs (seeds 0 and 2, position kp 6.2 and 5.0, never settling inside 5 cm). Four of the 12 wide designs carry the slow tail, against two of lag75's and one each of lag39's geometries.
  • Certificate and training. All 12 designs pass the margin test under the three identified laws (AA4). Rejected share 0.15, 0.13 against lag75's 0.15, 0.16 (AA6 refuted: the wider interval is not harder to certify). Mass 1.451, 1.422 kg on a 0.086, 0.092 m propeller, position kp 17.6, 16.6: the same small-propeller, loop-robust region 7z found, not 7x's heavier one.
  • Verdicts. AA1 held (27, 38 against midpoints 99.5, 105.5). AA2 held (against 36, 51). AA3 held by count, not separated. AA4 held (12 of 12). AA5 refuted. AA6 refuted. Checks: 7u's lag39, 7x's lag60 and 7z's lag75 reproduce their stress counts and 7y's and 7z's responses exactly.
  • Reading (fixed before the run): AA1 held says the slow-floor result of 7u-7z is mostly a coverage result: train across the whole plausible lag range anchored at the identified lag rather than at a slow nominal one. AA5 refuted removes the hoped-for bonus: the wide tune is not faster than the 75 ms tune, so the two are interchangeable on these tests and the choice is about what a builder can state: "my motors lie somewhere in 39-150 ms" is a statement about uncertainty, "tune for 75 ms" is a nominal value no identification supports. The finding to carry forward: the lag39 tune trained on exactly the identified law's flown interval (39.1-78.2 ms) crashes 72, 73 times at that law, while the wide tune, trained mostly on slower plants than it is tested on, crashes 11, 15. Matching the training distribution to the test distribution was the worst choice here; the search needs plants beyond the test range to learn margin against its edge.
  • Limits: one arm; one ceiling (150 ms, not sourced; 7z's); the floor/ceiling split is not a full factorial (no 75 ms floor with a 78.2 ms-wide interval); six seeds, one evaluation seed; response tests are calm air at the design lever.
  • Next: the ceiling sweep at the identified floor (39.1 ms × U(1, c) for c giving 100 and 200 ms) to find how much coverage is enough; a position-kp floor as a training constraint, since the slow tail is now in 4 of 12 designs; the same coverage test on the other uncertain inputs (torque scale, mass), where the stress test's plant range is still U(0.7, 1) and U(0.9, 1.3).

2026-11-18 · Milestone 7z: tune at a 75 ms floor and score it on both the stress test and the response test; does the frontier bend? (predictions written and committed before experiments/m7z_floor75/run.py was run at full budget; outcome below)

  • Why: milestone 7y's first next step. 7u-7x found every slower tuning floor (39.1, 45, 52, 60 ms D<sup>0</sup>) crashes less in the held-out stress test (lever 2, 45°, cap) at every identified lag, with no knee by 60 ms; 7y put the price in a builder's units: each step slower costs about 0.1 s of rise time on a 1 m step, overshoot falls, and wind drift grows by millimetres. Along the four floors each step bought 30-50 fewer stress crashes of 1152 (identified total, per geometry: 170, 117, 75, 36 above; 178, 119, 87, 51 below). The frontier has no knee yet; 75 ms is the next bracket.
  • Plan (experiments/m7z_floor75/run.py): 7x's lever-aware arm exactly (closure src_li, 45° tilt limit, A_P cap, wide kp box, integrator feed-forward, margin constraint, TWR 2, 3200 CMA-ES evaluations, 6 seeds × 2 starts × 2 geometries; 7x's patch) with the floor at 75 ms D<sup>0</sup> (training plants 75 ms × U(1, 2)) and certificate corners at 33, 75 and 150 ms. 75 ms is not sourced: it is a tuning floor, 25% beyond 60 ms and 92% beyond the slower identification. Scored on (a) 7x's held-out stress set (seed 9999, 64 episodes, four capped conditions, margin test) under 7x's seven laws plus 75 ms, with 7x's lag45, lag52 and lag60 designs alongside; (b) 7y's response tests (1 m step, 3 and 6 m/s calm-to-wind steps; lever 1, 45°, no turbulence or yaw steps, seed 9999, 64 episodes) under the three identified laws, lag60 and lag75.
  • Looked at before writing these: a smoke run (64 evaluations, 2 seeds, 8 episodes on seed 4242, not seed 9999) to check the script runs and the corners move to 33/75/150 ms (ρ at x = 0.5 is 1.0302 with and without the 33 ms corners). The smoke designs are unconverged (most candidates rejected) and say nothing about the questions.
  • Predictions (L2-45 = diverged flights at lever 2, 45°, cap, summed over 6 seeds, of 384; "identified total" = L2-45 summed over 33, 36 and 39.1 ms, of 1152; lag60 values in brackets, above, below; response statistics are 7y's: per-set median over seeds of each design's median over episodes, at 36 ms): - Z1 (prior 55%): slower still pays: lag75's identified total is below lag60's (36, 51) in both geometries. - Z2 (prior 40%): the knee shows at the fast end: at 33 ms lag75 crashes on more L2-45 flights than lag60 (8, 12) in both geometries. - Z3 (prior 70%): the certificate holds: at least 11 of 12 lag75 designs pass the margin test under all three identified laws. - Z4 (prior 65%): mass keeps rising: mean mass lag75 > lag60 (1.508, 1.474 kg) in both geometries. - Z5 (prior 70%): the rise keeps slowing: median rise time to 90% of the 1 m step, lag75 > lag60 (1.74, 1.77 s) in both geometries. - Z6 (prior 50%): the price stays small: lag75's rise time is at most 15% longer than lag60's in both geometries. - Z7 (prior 55%): overshoot keeps falling: lag75's step overshoot is at most lag60's (3.2, 2.7 cm) in both geometries.
  • Reading fixed in advance: Z1 held with Z6 held says slower still pays at 75 ms at a small price, so the frontier has not bent and the floor is a stated rise-time requirement, not a property of the crash count; Z1 held with Z6 refuted says the frontier bends: crashes still fall but the price per step grows, so the recommendation is the floor whose rise time meets the stated requirement (7y: about 0.1 s per step until now); Z1 refuted puts the knee between 60 and 75 ms and 60 ms is the tune. Z3 refuted says the certificate finally binds at a slow enough floor.
  • Analysis: counts per set, geometry and law; margin validity; rejected share; mass, propeller, kp; response metrics per set and law. Post hoc only: Mann-Whitney U per cell, lag75 against lag60. Checks: 7x's designs reproduce 7x's counts and 7y's lag60 response. Six seeds, one evaluation seed.

### Outcome (experiments/m7z_floor75/results.txt, verdicts.txt, posthoc.txt; 24 CMA-ES runs of 3200 evaluations, 3799 s on 4 cores, then 48 designs × 8 laws × 4 conditions × 64 episodes, 402 s, and the response tests, about 60 s; figure m7z_floor75.png; simulated)

  • Slower still pays inside the identified range, but the step from 60 to 75 ms buys little there. Identified total (L2-45 at 33 + 36 + 39.1 ms, of 1152): 29, 33 against 60 ms's 36, 51 (above, below; −19%, −35%). Along the five floors: 170, 117, 75, 36, 29 (above) and 178, 119, 87, 51, 33 (below). Per cell, post hoc Mann-Whitney on the six per-seed counts, p = 0.04-1.0: only 39.1 ms below (12 against 23) is separated; at 33 ms the two tunes tie (8 against 8, 10 against 12; Z2 refuted, no cost at the fast end).
  • Where the 75 ms tune clearly wins is at slower motors than any identified. At 60 ms flown lag 49, 45 against 105, 83 (p = 0.026, 0.15); at 75 ms 117, 103 against 193, 188 (p = 0.024, 0.010). A builder whose motors may be slower than the two published identifications gets the most from it.
  • The certificate holds (Z3): all 12 designs pass the margin test under the three identified laws (60 ms tune also 12). The training rejected share is 0.15, 0.16 (60 ms: 0.15, 0.11).
  • Not heavier: lighter, on a smaller propeller (Z4 refuted): 1.431, 1.435 kg against 1.508, 1.474, propeller 0.082, 0.092 m against 0.115, 0.107. So 7x's "slower buys robustness with mass" does not continue; the 75 ms search found robustness through the loop instead (position kp 19.1, 15.5 against 18.1, 17.9). Median tracking cost at lever 1, 45°, 36 ms is about 7% worse (−2.05, −1.90 against −1.90, −1.78).
  • The response price grows and the overshoot trend reverses. Rise time to 90% of a 1 m step 1.98, 1.86 s against 1.74, 1.77 s (+14%, +5%; Z5 held, Z6 held just inside 15%; per-seed p = 0.82, 0.24). Overshoot 4.8, 3.2 cm against 3.2, 2.7 (Z7 refuted: 7y's monotone fall stops at 60 ms). Settle time to 5 cm (unsettled flights counted as never) 3.22, 2.17 s against 2.37, 1.95: worse in both geometries, by 36% above. Peak drift after a 6 m/s wind step 3.3, 2.2 cm against 2.8, 2.4. Nothing diverged. Two above designs (seeds 3, 5; position kp 5.4, 7.5) never settle inside 5 cm, the same slow-tail pattern as 7y.
  • Crashes avoided per extra 0.1 s of rise (figure, right): step 52→60 ms 41 (above), 22 (below); step 60→75 ms 2.8 and 20. With the centre of pressure above, the frontier flattens past 60 ms; below, it keeps its slope.
  • Verdicts. Z1 held (29, 33 against 36, 51). Z2 refuted. Z3 held (12 of 12). Z4 refuted. Z5 held. Z6 held (ratios 1.14, 1.05). Z7 refuted. Checks: 7x's 60 ms designs reproduce 7x's counts and 7y's response exactly.
  • Reading (fixed before the run): Z1 held with Z6 held says slower still pays at 75 ms at a small price, so the frontier has not bent and the floor is a stated rise-time requirement, not a property of the crash count. The data qualifies it: inside the identified range the 60→75 ms step is not separated at six seeds and costs settle time and overshoot, so for motors near the two published identifications 60 ms stays the recommendation; 75 ms is the choice when the motor may be slower than identified (60-75 ms flown lags), where its gain is measured.
  • Process note: a first pass of verdicts.py read 9 of 12 certified for both 60 and 75 ms, because run.py rounded results.json to five significant digits and certified designs sit at ρ = 0.99999x against a strict ρ < 1 test. Caught against results.txt (computed before rounding), run.py changed to keep the stress results at full precision, scoring rerun from the cached training; every count in results.txt was identical.
  • Limits: one arm; the floor and ceiling (U(1, 2)) are still not separated, so the 75 ms tune trains up to 150 ms; 75 ms is not sourced; six seeds, one evaluation seed; the response tests are calm air at the design lever; the 5 cm band and 6 m/s wind are stated requirements.
  • Next: separate floor from ceiling (39.1 ms × U(1, 3.8), the same 150 ms ceiling); a position-kp or settle-time floor as a training constraint to remove the slow tail, now seen at every floor; 12 seeds for 75 against 60 ms inside the identified range, where six seeds do not separate them.

2026-11-18 · Milestone 7y: what does tuning slower cost at lever 1? A 1 m position step and a calm-to-wind step on 7x's designs (predictions written and committed before experiments/m7y_response/run.py was run at full size; outcome below)

  • Why: milestone 7x's third next step. 7u-7x found that tuning the lever-aware arm at a slower lag floor cuts crashes in the held-out stress test (lever 2, 45°) at every identified lag, with no knee up to 60 ms; the price showed as mass (+4-8%), a softer position loop (kp 18.1, 17.9 against 22.5, 21.9 1/s²) and about 5% worse median tracking cost at lever 1. The stress test rewards a slow, heavy vehicle by construction, and the lever-1 tracking cost mixes turbulence, yaw steps and a wind step into one number. Before a 75 ms floor or a requirement-constrained search, the price has to be measured in units a builder writes into a requirement: how long a position step takes to settle, how far it overshoots, and how far the vehicle drifts when the wind steps.
  • Plan (experiments/m7y_response/run.py): no retraining. 7x's 48 lever-aware designs (lag39 from 7u, lag45 from 7w, lag52 and lag60 from 7x; 6 seeds × 2 geometries) flown under the three identified laws (33, 36, 39.1 ms D<sup>0</sup>) at lever 1, 45° tilt limit, A_P cap, integrator feed-forward, the evaluation's plant randomisation and sensor noise, seed 9999, 64 episodes of 6 s, with turbulence and yaw steps switched off so the response is the loop's own. Tests: step (start 1 m off the hold point, still air; settle time = time after which the distance stays within 5 cm, overshoot past the hold point, time to 90%); wind 3 and wind 6 (start on the hold point in calm air; the mean wind steps to 3 m/s, the held-out set's wind-step size, or 6 m/s, along x at 1 s; peak drift, RMS over the 5 s after the step, mean offset over the last second). poshold.rollout gains an inert trace argument to return the position history.
  • Looked at before writing these: a smoke run (2 seeds per set, 8 episodes on seed 4242, not seed 9999) to check the script runs and the metrics are finite. The 5 cm band, the 1 m step and the 6 m/s wind are not sourced: they are a stated requirement of the kind a survey or inspection mission writes (the inherited GNSS-class noise is 2 cm), and the 3 m/s step is the held-out set's.
  • Predictions (cell = law × geometry, 6 cells; per-set statistic = median over seeds of each design's median over episodes): - Y1 (prior 70%): the softer loop settles slower: lag60's step settle time is longer than lag45's in at least 5 of 6 cells. - Y2 (prior 55%): the price is small: lag60's settle time is at most 25% longer than lag45's in every cell. - Y3 (prior 60%): the softer loop drifts further: lag60's peak drift after the 6 m/s wind step is larger than lag45's in at least 5 of 6 cells. - Y4 (prior 50%): the heavier, softer vehicle overshoots less: lag60's step overshoot is smaller than lag45's in at least 5 of 6 cells. - Y5 (prior 75%): nothing diverges at lever 1 in calm air: no flight of any design diverges in the step or the 3 m/s wind test under the identified laws. - Y6 (prior 55%): the four floors are ordered: median settle time rises monotonically lag39 ≤ lag45 ≤ lag52 ≤ lag60 in at least 4 of 6 cells.
  • Reading fixed in advance: Y1 and Y3 held with Y2 held says a 60 ms tune costs a builder little at lever 1 for 63% fewer stress crashes, so the 60 ms floor is the recommendation and a 75 ms floor is the next bracket. Y1 or Y3 held with Y2 refuted says the trade is real: the floor is set by the settle or drift requirement, and the next step is a search with that requirement as a constraint. Y1 and Y3 both refuted says the softer kp is not what sets the response (the cap and the derivative gain do), and 7x's tracking cost came from the turbulence, not the loop.
  • Analysis: per cell, the per-set medians; per-seed values; unsettled and diverged counts. Post hoc only: Mann-Whitney U per cell, lag60 against lag45, on the six per-seed medians. Six seeds, one evaluation seed.

### Outcome (experiments/m7y_response/results.txt, verdicts.txt, posthoc.txt; 48 designs × 3 laws × 3 tests × 64 episodes, 137 s on 4 cores; figure m7y_response.png; simulated)

  • In calm air at lever 1 the motor lag does not matter. Every metric is the same to three decimals under the 33, 36 and 39.1 ms laws, so the six cells are two (one per geometry) repeated three times. The lag only shows once the gust torque is near the torque limit, as in the lever-2 stress test. Nothing diverged in any test (Y5 held).
  • The slower tune costs rise time and buys overshoot, both monotone in the floor. Rise time to 90% of a 1 m step (above, below): 1.44, 1.39 s (39.1 ms tune), 1.58, 1.47 (45 ms), 1.64, 1.61 (52 ms), 1.74, 1.77 (60 ms); the 60 ms tune rises 10% and 20% slower than the 45 ms tune (post hoc Mann-Whitney p = 0.13, 0.015). Overshoot falls the other way: 6.4, 9.7 cm / 5.9, 7.6 / 3.8, 5.5 / 3.2, 2.7 cm, so the 60 ms tune overshoots half (above) to a third (below) as much as the 45 ms tune (Y4 held; p = 0.39, 0.041). Neither ordering was predicted; both are monotone in all six cells.
  • Settle time is not where the price is. The median time to stay within 5 cm is 2.04, 2.56 s (45 ms tune) against 2.37, 1.95 s (60 ms tune), with unsettled flights counted as later than 6 s: slower above (+16%), faster below (−24%), because the 60 ms tune's slower rise is paid back by its smaller overshoot. Y1 (slower in at least 5 of 6 cells) refuted at 3 of 6; Y2 (at most 25% slower) held; Y6 (settle time monotone in the floor) refuted, 0 of 6.
  • Wind drift grows by millimetres. Peak drift after a 6 m/s calm-to-wind step is 2.3, 1.9 cm (45 ms) against 2.8, 2.4 cm (60 ms): larger in every cell (Y3 held), but by 4-5 mm and not separated at six seeds (p = 0.31, 0.39). After a 3 m/s step it is about 1 cm for every set. In this model drag on a 1.4 kg quadrotor at 6 m/s is small next to the tilt the loop can command.
  • A slow tail in one or two designs per set, whatever the floor (post hoc). 8 of the 48 designs never settle inside 5 cm in 6 s on most flights: one each of lag39, lag45 and lag52 below, one of lag52 above, and three of lag60 (seeds 4 and 5 above, 5 below). They park 5-8 cm short of the hold point with little oscillation (peak-to-peak 0.5-2 cm over the last second), and all have position kp 4.5-9.7 1/s² against a set mean near 20: a slow approach, not a limit cycle. The lever-2 stress objective never asks for a fast final approach, so the search leaves some designs with one.
  • Verdicts. Y1 refuted (3 of 6). Y2 held (ratios 1.16, 0.76). Y3 held (6 of 6, by millimetres). Y4 held (6 of 6). Y5 held (0 diverged). Y6 refuted (0 of 6).
  • Reading (fixed before the run): Y3 held with Y2 held says a 60 ms tune costs a builder little at lever 1 for 63% fewer stress crashes, so the 60 ms floor is the recommendation and a 75 ms floor is the next bracket. With the qualifier the data adds: the price is a 10-20% slower rise and a few millimetres of wind drift, and the 60 ms tune also overshoots less. Read along the four floors (figure, left), each step slower trades about 0.1 s of rise time for 30-50 fewer stress crashes of 1152.
  • Limits: calm-air tests with no turbulence and no yaw steps, so they isolate the loop and say nothing about gust rejection near the torque limit; one 1 m step on one axis; the 5 cm band and the 6 m/s wind are a stated requirement, not sourced; the drag model is the inherited CD_A = 1 on the sized area (ASSUMED); per-set medians over six seeds, and the unsettled designs make settle time a censored statistic; one evaluation seed.
  • Next: a 75 ms floor scored on both 7x's stress test and this response test, so the frontier in the figure gets its next point; a settle-time or position-kp floor as a training constraint (it would remove the slow-tail designs, at a cost to measure); separate floor from ceiling (39.1 ms × U(1, 3)).

2026-11-17 · Milestone 7x: tune at 52 ms and 60 ms with the certificate also required at 33 ms; where does tuning slower stop paying? (predictions written and committed before experiments/m7x_knee/run.py was run at full budget; outcome below)

  • Why: milestone 7w's first next step. 7w found that tuning at an unsourced 45 ms floor beats 7u's 39.1 ms tune in 19 of 20 cells, with no cost yet at the fast identified lag (33 ms), so the knee lies beyond 45 ms. Its cost showed at the certificate: two lever-aware above designs sat on the margin boundary (ρ 1.0000-1.0002) at 33 and 36 ms, because nothing in training checked them there.
  • Plan (experiments/m7x_knee/run.py): 7w's lever-aware arm exactly (closure src_li solved exactly, 45° tilt limit, A_P cap, wide kp box, integrator feed-forward, margin constraint, TWR 2, 3200 CMA-ES evaluations, 6 seeds × 2 starts × 2 geometries) at two floors, 52 ms and 60 ms D<sup>0</sup> (training plants floor × U(1, 2)), with the certificate's lag corners at 33 ms, the floor and twice the floor, so the search only accepts designs certified at the fast identified lag. Only the lever-aware arm: 7i's finding (lever-aware crashes less than 7l at the same lag) has held five times, and one arm buys two floors for 7w's compute. 52 and 60 ms are not sourced: they are tuning floors chosen to bracket a knee (33% and 53% beyond the slower identification). Held-out exactly as 7u-7w (seed 9999, 64 episodes, four capped conditions, unpatched plant draw and margin test) under 7u's five laws plus 52 and 60 ms. Designs scored: the 24 new ones (lag52, lag60), 7w's 12 lever-aware 45 ms designs (lag45) and 7u's 12 lever-aware 39.1 ms designs (lag39).
  • Looked at before writing these: a smoke run (64 evaluations, 2 seeds, 8 episodes on seed 4242, not the held-out seed) to check the script runs and the corners move to 33/52/104 and 33/60/120 ms. At x = 0.5 the 33 ms corners do not change the margin radius (1.0304 with and without), so whether they ever bind is part of the question. The smoke designs (63 of 64 candidates rejected in some runs) are unconverged and say nothing about the questions.
  • Predictions (L2-45 = diverged flights at lever 2, 45°, cap, summed over 6 seeds, of 384; lag45 counts in brackets as above, below; "identified total" = L2-45 summed over 33, 36 and 39.1 ms, lag45 117, 119): - X1 (prior 55%): the knee is below 60 ms: at 39.1 ms lag60 crashes on more L2-45 flights than lag45 (52, 46) in both geometries. - X2 (prior 45%): 52 ms still pays: lag52's identified total is at most lag45's (117, 119) in both geometries. - X3 (prior 60%): slower costs first at the fast end: at 33 ms lag60 crashes on more L2-45 flights than lag45 (27, 34) in both geometries. - X4 (prior 75%): the 33 ms corner holds the certificate: at least 11 of 12 lag52 and at least 11 of 12 lag60 designs pass the margin test under all three identified laws (lag45: 10 of 12). - X5 (prior 60%): slower floors are harder to certify in training: mean rejected share of training candidates lag60 > lag52 > lag45 (0.12, 0.11) in both geometries. - X6 (prior 65%): tuning slower buys robustness with mass: mean mass lag60 > lag45 (1.450, 1.359 kg) in both geometries. - X7 (prior 65%): each tune pays at its own floor: at 60 ms lag60 crashes on fewer L2-45 flights than lag45 in both geometries.
  • Reading fixed in advance: X1 held with X2 held puts the knee between 52 and 60 ms and the recommendation at about 52 ms with the 33 ms corner; X1 held with X2 refuted puts it between 45 and 52 ms, so 45 ms (7w) is the tune; X1 refuted says slower still pays at 60 ms and the binding constraint is the certificate or mass, not the crash count, and the next bracket is 75 ms. X4 refuted says the fast corner does not protect the identified range and the certificate has to be a training objective, not only a constraint.
  • Analysis: counts per set, geometry and law; margin validity and the largest ρ per law; rejected share; mass, propeller, kp. Post hoc only: Mann-Whitney U per cell against lag45. Six seeds, one held-out set.

### Outcome (experiments/m7x_knee/results.txt, verdicts.txt, posthoc.txt, corner_bind.txt; 48 CMA-ES runs of 3200 evaluations, 9090 s on 4 cores, then 48 designs × 7 laws × 4 conditions × 64 episodes, 411 s; figure m7x_knee.png; simulated)

  • Still no knee at 60 ms: every slower floor crashes less at every lag, the fast end included. At lever 2, 45°, cap, summed over 6 seeds (above, below): at 33 ms 37, 46 (39.1 ms tune), 27, 34 (45 ms), 22, 24 (52 ms), 8, 12 (60 ms); at 39.1 ms 72, 73 / 52, 46 / 30, 31 / 17, 23. Over the identified range (33 + 36 + 39.1 ms, both geometries) the four floors total 348, 236, 162 and 87 crashes of 2304 flights: each step slower cuts crashes by a third to a half (32%, 31%, 46%), and the 60 ms tune crashes 63% less than 7w's 45 ms tune and 75% less than 7u's slow-end tune. Across the four floors the ordering is monotone in 13 of 14 law × geometry cells, the exception (inherited 15 ms law, below: 52 ms 10 against 45 ms 8).
  • This one is a measured gap. Post hoc, the 60 ms designs against the 45 ms designs: Mann-Whitney p = 0.004-0.014 in all six identified cells; the 52 ms designs against 45 ms are not separated inside the identified range (p 0.14-0.87).
  • The certificate holds everywhere. All 24 new designs pass the margin test under all three identified laws (X4; lag45 10 of 12), and 23 of 24 under the inherited 15 ms law. The 33 ms corner is not why (post hoc, corner_bind.txt): on 22 of 24 designs the certificate radius is the same with only the floor corners or only the 33 ms corners, so the binding mode does not move with the lag; the 33 ms corner alone sets it on one design. 7x differs from 7w in effect only by the floor.
  • The price is mass and gain, not crashes. The 60 ms designs are 4% and 8% heavier than the 45 ms designs (1.508, 1.474 kg against 1.450, 1.359), with a larger propeller (0.115, 0.107 m against 0.103, 0.100) and a softer position loop (kp 18.1, 17.9 against 22.5, 21.9); median tracking cost at lever 1, 45°, 36 ms is about 5% worse (−1.90, −1.78 against −1.81, −1.70). The 52 ms designs carry no mass cost (1.411, 1.361 kg) and lose about 3% in tracking.
  • Verdicts. X1 refuted (60 ms crashes less at 39.1 ms in both geometries: 17, 23 against 52, 46). X2 held (identified totals 75, 87 against 117, 119). X3 refuted (8, 12 against 27, 34 at 33 ms). X4 held (12 and 12 of 12). X5 refuted (rejected share 0.15 > 0.14 > 0.12 above, but 0.11 < 0.12 below). X6 held. X7 held (105, 83 against 222, 224 at 60 ms). Check: 7w's designs reproduce 7w's counts exactly.
  • Reading (fixed before the run): X1 refuted says slower still pays at 60 ms, and the binding constraint is the certificate or mass, not the crash count; the next bracket is 75 ms. The certificate did not bind; mass and tracking began to move at 60 ms. So in this simulator the lever-2 crash count alone never sets the floor: a slower tune buys robustness with a heavier vehicle and a softer position loop, and where to stop is a tracking and endurance requirement, not a stability one.
  • Limits: one arm (lever-aware); the plant randomisation U(1, 2) multiplies the floor, so a 60 ms floor also trains up to 120 ms and floor and ceiling are still not separated; the held-out stress test (twice the trained lever, 45°) rewards a slow, heavy vehicle by construction, and nothing here charges for the softer position loop beyond the lever-1 tracking cost; 52 and 60 ms are not sourced; six seeds, one held-out set, first-order lag, two identifications from one lab.
  • Next: separate floor from ceiling (39.1 ms × U(1, 3), same 120 ms ceiling as here); a 75 ms floor with a tracking-cost or flight-time constraint, so the knee is set by a requirement a builder states; a gust-rejection or step-tracking test at lever 1 that charges for the softer position loop; 12 seeds for the 52 ms floor against 45 ms.

2026-11-17 · Milestone 7w: tune at 45 ms, beyond the slower identification; where does tuning slower stop paying? (predictions written and committed before experiments/m7w_slow45/run.py was run at full budget; outcome below)

  • Why: milestone 7v's second next step. 7u found tuning at 39.1 ms (the slower identification) beats the 36 ms midpoint tune in all 16 cells; 7v found the gain is the slow floor of the training plants, not the width of the draw, and that widening downward buys nothing at the fast end. If the rule is "train on plants no faster than the slow end", the question left is how far past the slow end it keeps paying. 7u's first warning sign is the certificate under the inherited 15 ms law (20 of 24, against 7t's 23).
  • Plan (experiments/m7w_slow45/run.py): 7u's two arms exactly (7l and lever-aware objectives, closure src_li solved exactly, 45° tilt limit, A_P cap, wide kp box, integrator feed-forward, margin constraint, TWR 2, 3200 CMA-ES evaluations, 6 seeds × 2 starts × 2 geometries) with only the tuning lag changed to 45 ms D<sup>0</sup>: training plants 45 ms × U(1, 2) (45-90 ms), certificate corners 45 and 90 ms. 45 ms is not sourced; it is 7u's stress value (15% beyond the slower identification), chosen on purpose as a floor above the identified range to bracket where slower stops paying. Held-out exactly as 7u and 7v (seed 9999, 64 episodes, four capped conditions, margin test, unpatched plant draw and corners) under the same five laws. Designs scored: the 24 new ones (lag45), 7u's 24 (lag39) and 7t's 24 (lag36).
  • Looked at before writing these: a smoke run (64 evaluations, 2 seeds, 8 episodes on seed 4242, not the held-out seed) to check the script runs and the certificate corners move to 45/90 ms. The smoke designs are unconverged and say nothing about the questions.
  • Predictions (L2-45 = diverged flights at lever 2, 45°, cap, summed over 6 seeds, of 384; lag39 counts in brackets, order 7l above, 7l below, lever-aware above, lever-aware below): - W1 (prior 60%): slower costs at the fast end: at 33 ms lag45 crashes on more L2-45 flights than lag39 (80, 107, 37, 46) in at least 3 of 4 arm-geometries. - W2 (prior 50%): slower still pays at the slow identification: at 39.1 ms lag45's L2-45 is at most lag39's (150, 167, 72, 73) in at least 3 of 4. - W3 (prior 60%): tuning at the stress value pays at the stress value: at 45 ms lag45 crashes on fewer L2-45 flights than lag39 (215, 225, 149, 141) in at least 3 of 4. - W4 (prior 50%): the identified-range certificate starts to go: at most 21 of 24 lag45 designs pass the margin test under all three identified laws (lag39 24 of 24). - W5 (prior 70%): at 45 ms at least 22 of 24 lag45 designs pass the margin test (lag39 15, lagrand 8). - W6 (prior 60%): the fast end of the certificate goes further: under the inherited 15 ms law at most 17 of 24 lag45 designs pass (lag39 20). - W7 (prior 65%): 7i's finding survives a fifth time: at 39.1 ms the lever-aware lag45 crashes on fewer L2-45 flights than 7l lag45 in both geometries and is at least 5% heavier in both.
  • Reading fixed in advance: W1 held with W2 refuted says 39.1 ms is at or past the knee: tune at the slow identification, not beyond it. W2 held says slower keeps paying inside the identified range and the knee lies beyond 39.1 ms; a 52 ms tune is the next bracket. W1 refuted with W2 held says the error is one-sided across this whole span and the tuning lag should be the slowest a margin certificate still covers.
  • Analysis: counts per set, geometry, law and condition; margin validity; mass, propeller, kp. Post hoc only: Mann-Whitney U per cell, lag45 against lag39. Six seeds, one held-out set.

### Outcome (experiments/m7w_slow45/results.txt, verdicts.txt, posthoc.txt; 48 CMA-ES runs of 3200 evaluations, 6115 s on 4 cores, then 72 designs × 5 laws × 4 conditions × 64 episodes, 413 s; figure m7w_slow45.png; simulated)

  • Tuning slower still pays, and it pays everywhere, including at the fast end. At lever 2, 45°, cap, the 45 ms designs crash on fewer flights than 7u's 39.1 ms designs in 19 of 20 law × arm-geometry cells (one tie, 26 against 26 for 7l above under the inherited 15 ms law). Inside the identified range: at 33 ms 75, 77, 27, 34 against 80, 107, 37, 46; at 36 ms 88, 98, 38, 39 against 109, 132, 61, 59; at 39.1 ms 123, 108, 52, 46 against 150, 167, 72, 73 (7l above, 7l below, lever-aware above, lever-aware below). Summed over the four arm-geometries the identified-range totals are 213, 263, 329 against 270, 361, 462 (7t's midpoint tune: 341, 455, 603), 21-29% fewer crashes than the slow-end tune.
  • No knee yet. The prediction that slower would start to cost at 33 ms (W1) is refuted in all four cells. In this span the lag error stays one-sided: a tune for a slower motor than the one flown has not yet cost anything this test sees, and the knee lies beyond 45 ms.
  • The cost is at the certificate, and it sits on the boundary. 22 of 24 designs pass the margin test under all three identified laws (7u: 24). The two that fail are lever-aware above seeds 0 and 5, at ρ = 1.0000-1.0002 under 33 and 36 ms (the test is strict, ρ < 1): the search pushes these designs onto the certificate's boundary at the 45/90 ms corners and they stay on it, not clearly outside it, at faster lags. Under the inherited 15 ms law 20 of 24 pass, the same as 7u (four failures, all ρ ≤ 1.0004). At 45 ms all 24 are certified (7u 15, 7v 8).
  • Same vehicles, no tracking cost. Mass 1.09-1.45 kg against 1.05-1.42; propeller diameter the same or smaller (0.100-0.119 m against 0.103-0.137 m); position kp 21.9-23.1 against 20.8-25.0. Median J at lever 1, 45°, 36 ms is no worse (−1.95, −1.66, −1.81, −1.70 against −2.41, −2.23, −1.83, −1.67).
  • How sure. Every arm-geometry agrees at every identified law and at 45 ms (sign test two-sided p = 0.125 each, the floor for four cells); per cell Mann-Whitney p 0.05-0.47 inside the identified range, 0.045-0.42 at 45 ms. Read: tuning at 45 ms probably beats tuning at 39.1 ms across the whole identified range; not a measured gap at six seeds.
  • 7i's finding holds a fifth time: at 39.1 ms lever-aware 45 ms designs crash on 52 against 123 (above) and 46 against 108 (below) and are 9.2% and 24.8% heavier.
  • Verdicts. W1 refuted (0 of 4). W2 held (4 of 4). W3 held (4 of 4). W4 refuted (22 of 24; needed at most 21). W5 held (24 of 24). W6 refuted (20 of 24; needed at most 17). W7 held. Check: 7u's designs reproduce 7u's counts at 33 and 39.1 ms exactly.
  • Reading (fixed before the run): W1 refuted with W2 held says the error is one-sided across this span and the tuning lag should be the slowest a margin certificate still covers. Here the certificate is the binding constraint: at 45 ms it is already on the boundary for two of 24 designs at the fast identified lags. 45 ms is not sourced; it is a tuning choice, not a claim about any motor.
  • Limits: one slower floor; the plant randomisation U(1, 2) multiplies it, so 45 ms training plants span 45-90 ms against 39.1-78.2 ms, and the two effects (slower floor, higher ceiling) are not separated. Six seeds, one held-out set, first-order lag only, two identifications from one lab.
  • Next: 52 ms and 60 ms floors to find the knee, with the certificate also checked at the fast identified corner (33 ms) as a constraint, so a slower tune cannot trade away the identified range; separate the floor from the ceiling (train 39.1 ms × U(1, 2.3), the same 90 ms ceiling as here); 12 seeds for the best floor against 7u.

2026-11-17 · Milestone 7v: train across the identified lag (33-39.1 ms) instead of at its slow end; is 7u's gain the slow value or the width? (predictions written and committed before experiments/m7v_lag_rand/run.py was run at full budget; outcome below)

  • Why: milestone 7u's first next step. 7u found that tuning at 39.1 ms (the slower of two identifications) beats 7t's 36 ms midpoint tune in all 16 law × arm-geometry cells and keeps the certificate across 33-39.1 ms, but loses the certificate on 4 of 6 lever-aware above designs under the inherited 15 ms law. It left open whether the gain comes from the slow value or from conservatism the search finds by other routes.
  • What a "tuning lag" is in these arms (read from poshold.rollout and linstab, written down here because earlier entries call it the "nominal model"): with the integrator feed-forward (ff_mode "integ") the controller holds no internal lag model; the lag enters only as the floor of the training plants (vehicle lag × plant randomisation U(1, 2)) and as the certificate's plant corners (lag × {1, 2}). A 39.1 ms tune therefore never trains on, or certifies at, a plant faster than 39.1 ms; the 33 ms held-out flights are the first it meets.
  • Plan (experiments/m7v_lag_rand/run.py): 7u's two arms exactly (7l and lever-aware objectives, closure src_li solved exactly, 45° tilt limit, A_P cap, wide kp box, integrator feed-forward, margin constraint, TWR 2, 3200 CMA-ES evaluations, 6 seeds × 2 starts × 2 geometries) with only the training lag changed: per episode 33 ms × U(1, 39.1/33) × U(1, 2), so the identified range times 7u's plant randomisation (33-78.2 ms), and the certificate's lag corners at 33 and 78.2 ms (both ends of the range certified; 7u's were 39.1 and 78.2). Held-out exactly as 7u (seed 9999, 64 episodes, four capped conditions, margin test, unpatched plant draw and corners) under 7u's five laws: 33, 36, 39.1 ms D<sup>0</sup>, the unsourced 45 ms stress value, the inherited 15 ms D<sup>1.5</sup>. Designs scored: the 24 new ones (lagrand), 7u's 24 (lag39) and 7t's 24 (lag36).
  • Looked at before writing these: a smoke run (64 evaluations, 2 seeds, 8 episodes on seed 4242, not the held-out seed) to check the script runs and that the patch reaches training (certificate corners 33/78.2 ms against 39.1/78.2; training lags 34.9-73.9 ms on 8 draws). At x = 0.5 the margin radius is the same under both corner sets (1.0301), so the slow corner binds there; whether the fast corner ever binds is part of the question. The smoke designs are unconverged and say nothing about the questions.
  • Predictions (L2-45 = diverged flights at lever 2, 45°, cap, summed over 6 seeds, of 384; 7u's lag39 counts in brackets, order 7l above, 7l below, lever-aware above, lever-aware below): - V1 (prior 80%): at least 22 of 24 lagrand designs pass the margin test under all three identified laws. - V2 (prior 55%): the gain is the slow floor, so widening down costs at the slow end: at 39.1 ms lagrand's L2-45 is at least lag39's (150, 167, 72, 73) in at least 3 of 4 arm-geometries. - V3 (prior 50%): training on 33 ms plants pays at 33 ms: lagrand crashes on fewer L2-45 flights than lag39 (80, 107, 37, 46) in at least 3 of 4. - V4 (prior 60%): the cost is small either way: lagrand's worst identified L2-45 is at most 1.15 × lag39's (150, 167, 72, 73) in all 4. - V5 (prior 55%): certifying the 33 ms corner carries to the fast end: under the inherited 15 ms law at least 22 of 24 lagrand designs pass the margin test (lag39: 20 of 24). - V6 (prior 45%): beyond the range nothing is gained: at 45 ms at most 15 of 24 lagrand designs pass the margin test (lag39: 15). - V7 (prior 65%): 7i's finding survives again: at 39.1 ms the lever-aware lagrand crashes on fewer L2-45 flights than 7l lagrand in both geometries and is at least 5% heavier in both.
  • Reading fixed in advance: V2 held and V3 held is a trade (the range tune buys the fast end with the slow end); V2 refuted (the range tune no worse at 39.1 ms) says 7u's gain was conservatism from the slow plants, which the wider draw keeps; V3 refuted with V2 held says tune at the slow end and stop.
  • Analysis: counts per set, geometry, law and condition; margin validity; mass, propeller, kp. Post hoc only: Mann-Whitney U per cell, lagrand against lag39. Six seeds, one held-out set.

### Outcome (experiments/m7v_lag_rand/results.txt, verdicts.txt, posthoc.txt; 48 CMA-ES runs of 3200 evaluations, 5360 s on 4 cores, then 72 designs × 5 laws × 4 conditions × 64 episodes, 365 s; figure m7v_lag_rand.png; simulated)

  • Training across the range is worse than training at its slow end in 11 of 12 identified cells, including every cell at the fast end it added. At lever 2, 45°, cap, the range-trained designs crash on more flights than 7u's 39.1 ms designs at both ends of the range in all four arm-geometries: at 33 ms 101, 112, 61, 53 against 80, 107, 37, 46; at 36 ms 144, 129, 90, 60 against 109, 132, 61, 59 (the exception, 7l below, 129 against 132); at 39.1 ms 191, 177, 132, 90 against 150, 167, 72, 73 (7l above, 7l below, lever-aware above, lever-aware below). Seeing 33 ms plants in training did not buy anything at 33 ms.
  • So 7u's gain is the slow floor of the training plants, not the width of the draw. Adding faster plants to the training draw pulls the designs back toward 7t's midpoint tune (post hoc, not predicted): inside the identified range the range-trained designs sit between the two, fewer crashes than 7t's 36 ms designs in 9 of 12 law × arm-geometry cells (the exception every time is lever-aware above), more than 7u's in 11 of 12. The training draw's mean lag is 36 ms, the midpoint 7t tuned at.
  • Same vehicles, softer gains elsewhere. Mass, propeller and position kp barely move (1.09-1.42 kg against 1.05-1.42; kp 21.1-22.7 against 20.8-25.0), and tracking at 36 ms is no better (median J at lever 1, 45°: −3.96, −2.35, −2.12, −1.66 against −2.41, −2.23, −1.83, −1.67), so the difference sits in the attitude gains the search settles on when some training plants are fast.
  • The 33 ms certificate corner does not buy the fast end either. All 24 range-trained designs pass the margin test across 33-39.1 ms (as do 7u's), and under the inherited 15 ms law 21 of 24 pass against 20 of 24 (V5 needed 22). Beyond the range it costs: at 45 ms 8 of 24 stay certified against 15 of 24.
  • How sure. Every identified cell leans the same way (sign test over the four arm-geometries, two-sided p = 0.125 at 33 and at 39.1 ms); per cell the largest gaps are in lever-aware above (Mann-Whitney p 0.05-0.09), the others 0.13-0.94. Read: no gain from the range draw, a probable loss of 5-83% more crashes per cell where it loses, not a measured one. Six seeds, one held-out set.
  • 7i's finding holds a fourth time: at 39.1 ms lever-aware range-trained designs crash on 132 against 191 (above) and 90 against 177 (below) and are 12.4% and 25.2% heavier.
  • Verdicts. V1 held (24 of 24). V2 held (4 of 4). V3 refuted (0 of 4: the range draw is worse at 33 ms too). V4 refuted (1 of 4 within 1.15×; ratios 1.06-1.83). V5 refuted (21 of 24). V6 held (8 of 24). V7 held. Check: 7u's designs reproduce 7u's counts at 33 and 39.1 ms exactly.
  • Reading (fixed before the run): V2 held with V3 refuted means tune at the slow end and stop. For this class of vehicle and controller a tuner that knows only a lag range should train on plants no faster than its slow end; widening the draw downwards gives up robustness and buys nothing at the fast end.
  • Limits: the range draw is uniform on 33-39.1 ms times U(1, 2); a draw weighted to the slow end, or a robust (worst-case) objective over the range, are not tried. Two identifications from one lab; first-order lag only; six seeds and one held-out set; the "pulled toward the midpoint" reading and the gain attribution are post hoc.
  • Next: 12 seeds for the slow-end arm against 7t's midpoint arm to put a number on the gap; a slower floor (45 ms) as a training law, to find where tuning slower stops paying (7u's 15 ms certificate loss is the first sign); the 7r next steps (flight-time requirement; LiPo at about 12 min) at 39.1 ms; a step-response measurement for one 4-inch and one 6-inch motor-propeller pair to source the diameter exponent.

2026-11-17 · Milestone 7u: tune at the slow end of the identified lag; does that alone cover the identified range? (predictions written and committed before experiments/m7u_slow_end/run.py was run at full budget; outcome below)

  • Why: milestone 7t's first next step. 7t retuned at 36 ms, the midpoint of the two identifications (33 ms NeuroBEM, 39.1 ms Agilicious), and found the lag error asymmetric: designs tuned for a slower lag than they fly lose little (its 36 ms designs crash less at 15 ms than the 15 ms tunes), designs tuned for a faster one fail outright. If that holds across the identified range, a tuner without an identified lag should tune at the slow end and pay nothing for it; if it does not, the range has to be trained or certified across.
  • Plan (experiments/m7u_slow_end/run.py): 7t's two arms exactly (7l and lever-aware objectives, closure src_li solved exactly, 45° tilt limit, A_P cap, wide kp box, integrator feed-forward, margin constraint, TWR 2, 3200 CMA-ES evaluations, 6 seeds × 2 starts × 2 geometries) with only the lag law changed to 39.1 ms for every propeller (the slower identification, D<sup>0</sup>); vehicle and nominal model together, plant randomisation (lag × 1-2) on top. Held-out as 7t (seed 9999, 64 episodes; lever 1 at 45°, lever 2 at 20° and 45°, lever 3 at 20°, all capped; margin test) under five laws: 33, 36 and 39.1 ms D<sup>0</sup>; a stress value of 45 ms D<sup>0</sup> (not sourced: 15% beyond the slower identification, standing in for an unidentified slower motor or a sagging pack); and the inherited 15 ms D<sup>1.5</sup>. Designs scored: the 24 new ones (lag39) and 7t's 24 (lag36).
  • Looked at before writing these: a smoke run (64 evaluations, 2 seeds, 8 episodes on seed 4242, not the held-out seed) to check the script runs; its lag39 designs are unconverged and say nothing about the questions. It also scored 7t's converged designs on those 8 episodes of seed 4242 (at 45 ms they lose the margin test in 3 of 4 smoke designs).
  • Predictions (L2-45 = diverged flights at lever 2, 45°, cap, summed over 6 seeds, of 384; "worst identified" = the largest L2-45 over the 33, 36 and 39.1 ms laws): - U1 (prior 85%): at least 22 of the 24 lag39 designs pass the margin test at 39.1 ms. - U2 (prior 60%): tuning for the lag pays at the slow end: at 39.1 ms, lag39 crashes on fewer L2-45 flights than lag36 (7t: 198, 195, 93, 117) in at least 3 of 4 arm-geometries. - U3 (prior 55%): the slow-end tune costs little at the midpoint: at 36 ms, lag39's L2-45 is at most 1.15 × lag36's (149, 161, 63, 82) in at least 3 of 4 arm-geometries. - U4 (prior 55%): minimax: lag39's worst identified L2-45 is below lag36's in at least 3 of 4 arm-geometries. - U5 (prior 60%): the certificate covers the range: at least 22 of 24 lag39 designs pass the margin test under all three identified laws, against at most 21 of 24 lag36 designs (7t: 21 of 24 at 39.1 ms). - U6 (prior 50%): beyond the range: at 45 ms at least 12 of 24 lag39 designs pass the margin test, and lag39 crashes on fewer L2-45 flights than lag36 in all 4 arm-geometries. - U7 (prior 65%): 7i's finding survives again: at 39.1 ms the lever-aware lag39 crashes on fewer L2-45 flights than 7l lag39 in both geometries and is at least 5% heavier in both. - U8 (prior 60%): the tuner slows the position loop further: mean position kp of lag39 is below lag36's in all 4 arm-geometries.
  • Analysis: counts per set, geometry, law and condition; margin validity; mass, propeller, kp; Spearman ρ(mass, L2-45) per law. No test beyond counts (six seeds, one held-out set).

### Outcome (experiments/m7u_slow_end/results.txt, verdicts.txt, posthoc.txt; 48 CMA-ES runs of 3200 evaluations, 5075 s on 4 cores, then 48 designs × 5 laws × 4 conditions × 64 episodes, 233 s; figure m7u_slow_end.png; simulated)

  • Tuning at the slow end costs nothing across the identified range, and the comparison leans its way everywhere. At lever 2, 45°, cap, the 39.1 ms designs crash on fewer flights than 7t's 36 ms designs under every identified law in all four arm-geometries: at 33 ms 80, 107, 37, 46 against 108, 124, 53, 56; at 36 ms (7t's own training law) 109, 132, 61, 59 against 149, 161, 63, 82; at 39.1 ms 150, 167, 72, 73 against 198, 195, 93, 117 (7l above, 7l below, lever-aware above, lever-aware below).
  • The certificate covers the range. All 24 slow-end designs pass the margin test at 33, 36 and 39.1 ms; 20 of 7t's 24 do. At the unsourced 45 ms stress value 15 of 24 slow-end designs stay certified against 10 of 24, and they still crash less (215, 225, 149, 141 against 248, 251, 168, 190).
  • How large the gap is, honestly. Every one of 16 law × arm-geometry cells favours the slow-end tunes, but with six designs per set no single cell is significant (post hoc Mann-Whitney U, two-sided, p 0.08-0.81). The four arm-geometries are independent design sets and all four agree under every law (one-sided sign test over four, p = 0.0625). Read: no cost, a probable gain of about 10-35% fewer crashes, not yet a measured one.
  • Why the midpoint tune lost at its own lag (post hoc). Not mass or propeller: the slow-end designs weigh the same (1.05-1.42 kg against 1.05-1.42) on similar propellers (0.103-0.137 m against 0.111-0.124 m). They track better at 36 ms too (median J at lever 1, 45°: −2.41, −2.23, −1.83, −1.67 against −4.61, −2.86, −1.98, −1.98), so tuning for a lag the controller is slightly pessimistic about acts as a robustness margin on the gains, consistent with 7t's one-sided finding. Training adds a plant lag randomisation of × 1-2 on the nominal law, so neither tune saw a plant faster than its own nominal value; the held-out flights at 33 ms are the only faster-than-nominal plants either set meets.
  • The price is at the far fast end. Under the inherited 15 ms D<sup>1.5</sup> law the counts tie (26, 60, 8, 13 against 29, 59, 8, 15), but 4 of 6 lever-aware above slow-end designs (propeller 0.103 m, lag 11 ms under that law) fail the margin test there; the 36 ms ones all pass. A design tuned for a slow motor loses its certificate on a much faster one, not its flights.
  • 7i's finding holds again: at 39.1 ms the lever-aware designs crash on 72 against 150 (above) and 73 against 167 (below) and are 8.6% and 28.8% heavier; ρ(mass, L2-45) is −0.77 to −0.95 under every law.
  • Verdicts. U1 held (24 of 24). U2 held (4 of 4). U3 held (4 of 4; ratios 0.72-0.97). U4 held (4 of 4). U5 held (24 against 20 of 24). U6 held (15 of 24 certified; fewer crashes in 4 of 4). U7 held. U8 refuted: the slow-end tunes do not soften the position loop further (kp 23.2, 25.0, 20.8, 22.0 against 25.3, 24.6, 23.7, 21.3). Check: 7t's designs reproduce 7t's counts at 36 and 39.1 ms exactly.
  • Reading: for this class of vehicle, a tuner that knows only the identified range (33-39 ms) should tune at its slow end: within the range that is never worse in these runs and is better in every cell, and it keeps the certificate across the range and more often beyond it. Randomising the lag over 33-39 ms in training is no longer needed to get coverage; it remains a check of whether the gain comes from the slow nominal value or from conservatism the CMA-ES search can find by other routes. The programme's tuning law from here is 39.1 ms, D<sup>0</sup>.
  • Limits: two identifications from one lab (2306 motors, 5-5.1 inch); first-order lag only; nominal model and vehicle share the lag (the only mismatch is the plant randomisation × 1-2 and the scoring laws); six seeds and one held-out set, so cell-level differences are within seed noise; the 45 ms stress value is not sourced; the analysis of why (tracking at 36 ms, the 15 ms certificate loss) is post hoc.
  • Next: a lag-randomised tune (33-39 ms in the nominal model, not only the plant) to tell a slow nominal value from conservatism; 12 seeds for the slow-end and midpoint arms if the gap is to be stated as a number; the 7r next steps (flight-time requirement instead of a fixed 15 min; LiPo at about 12 min) at 39.1 ms; a step-response measurement for one 4-inch and one 6-inch motor-propeller pair to source the diameter exponent.

2026-11-17 · Milestone 7t: retune at the motor lag two published identifications give (predictions written and committed before experiments/m7t_lag_src/run.py was run at full budget; outcome below)

  • Why: milestone 7s's first next step. 7s found that the programme's counts, its stability certificate and the retuned designs' parity with the published ones depend on a motor lag law, 0.015 (D / 0.127 m)^1.5 s, that was never sourced, and that at 30 ms almost no design passes the margin test.
  • Sourcing (data/motor_lag/identified.csv, lag_law.json; papers read 2026-11-17): two system identifications of brushless 5-5.1 inch, 2306-motor quadrotors of about 0.75 kg give 33 ms (rotor-speed time constant, Bauersfeld et al. 2021, NeuroBEM) and 39.1 ms (motor time constant from force-sensor fits, Foehn et al. 2022, Agilicious), 2.2-2.6 times the inherited 15 ms at 0.127 m. Rows read but not used for the fit, with the reason in use: the Crazyflie (brushed, 46 mm: 72 ms identified, about 73 ms from the manufacturer's step response, 65 ms quoted from Förster 2015), a 3.35 kg quadrotor with no propeller stated (65 ms in its figure, while its text calls it "much smaller" than the Crazyflie's), and simulator settings (RotorPy Hummingbird 5 ms, the Agilicious simulator 30 ms). The MiniQuadTestBench throttle-response traces would give step data per motor, but they are served only through the site's admin data endpoint, which this session did not use. Neither identification measures a diameter exponent.
  • Plan (experiments/m7t_lag_src/run.py): 7r's two arms exactly (7l src_li, aware src_li: closure src_li solved exactly, 45° tilt limit, A_P cap, wide kp box, integrator feed-forward, margin constraint, TWR 2, 3200 CMA-ES evaluations, 6 seeds × 2 starts × 2 geometries), with only the lag law changed to 36 ms for every propeller (the midpoint of the two identifications; exponent 0, since none is sourced and the design range 0.106-0.134 m sits next to the identified 0.127-0.130 m). The law sets the vehicle and the controller's nominal model together, as in 7s; the plant randomisation (lag × 1-2) stays on top. Held-out as 7s (seed 9999, 64 episodes; lever 1 at 45°, lever 2 at 20° and 45°, lever 3 at 20°, all capped; margin test) under five laws: 36 ms D^0 (training), 33 and 39.1 ms D^0 (the two identifications), 36 ms D^1.5, and the inherited 15 ms D^1.5. Designs scored: the 24 new ones and 7r's 24 retuned ones.
  • Looked at before writing these: a smoke run (64 evaluations, 2 seeds, 8 episodes on seed 4242, not the held-out seed) to check the script runs; its designs are unconverged and say nothing about the questions.
  • Predictions (L2-45 = diverged flights at lever 2, 45°, cap, summed over 6 seeds, of 384): - T1 (prior 70%): the tuners still find certified designs: at least 20 of the 24 new designs pass the margin test under the training law. - T2 (prior 75%): tuning for the lag pays: under the training law, each new arm-geometry crashes on fewer L2-45 flights than 7r's retuned designs of the same objective and geometry under the same law (4 of 4). - T3 (prior 60%): a slower actuator costs robustness even when tuned for: under the training law, the new designs crash on more L2-45 flights than 7r's retuned designs did under the inherited law (106 and 131 for 7l, 54 and 64 lever-aware) in at least 3 of 4 arm-geometries. - T4 (prior 55%): 7i's finding survives: under the training law, aware lag36 crashes on fewer L2-45 flights than 7l lag36 in both geometries. - T5 (prior 50%): it still does so by buying mass: aware lag36's mean mass is at least 5% above 7l lag36's in both geometries. - T6 (prior 55%): with no lag reward for a small propeller, the tuners stop shrinking it: mean propeller diameter of the new designs is at least 0.120 m in all 4 arm-geometries (7r: 0.106-0.115 m). - T7 (prior 60%): the new designs sit on their own stability boundary again (as 7c and 7s found): under the slower identification (39.1 ms) at most 12 of the 24 pass the margin test. - T8 (prior 70%): the tuners slow the position loop: mean position kp of the new designs is below 7r's retuned designs' in all 4 arm-geometries.
  • Analysis: counts per set, geometry, law and condition; margin validity; mass, propeller, kp; Spearman ρ(mass, L2-45) per law. No test beyond counts (six seeds, one held-out set).

### Outcome (experiments/m7t_lag_src/results.txt, verdicts.txt; 48 CMA-ES runs of 3200 evaluations, 6110 s on 4 cores, then 48 designs × 5 laws × 4 conditions × 64 episodes, 277 s; figure m7t_lag_src.png; simulated)

  • At the identified lag, the 15 ms tunes do not fly. Under the training law (36 ms) none of 7r's 24 retuned designs passes the margin test, and they crash on 374-383 of 384 lever-2, 45° flights; even at lever 1, 45° (the condition the 7l objective trains on) they lose 353-370 of 384. Under either identification alone (33 or 39.1 ms) the picture is the same (359-384 at L2-45, 0 of 24 certified). 7s's 30 ms bracket was close to the sourced value, and it was the bad end.
  • Retuned at 36 ms, the tuners find certified designs that fly. All 24 new designs pass the margin test; lever-2, 45° crashes are 149 and 161 (7l objective) and 63 and 82 (lever-aware) of 384, against 106 and 131 and 54 and 64 for 7r's designs at their own 15 ms law. The slower actuator costs 17-41% more crashes when tuned for, not the near-total loss of not tuning for it. Tracking at lever 1, 45° stays close to 7r's at its own law for the lever-aware arm (median J −1.98 against −1.54 to −1.52) and worsens for the 7l arm (−4.61 and −2.86 against −1.61 and −1.81).
  • 7i's finding survives the sourced lag, and still through mass. The lever-aware designs crash on 63 against 149 (above) and 82 against 161 (below) and are 14% and 28% heavier (1.42 against 1.24 kg; 1.34 against 1.05 kg). Over the 24 new designs per geometry, ρ(mass, L2-45) is −0.50 to −0.94 under all five scoring laws.
  • How the tuners pay for the lag: a softer position loop (mean kp 21-25 s⁻² against 27-31 for 7r) and more mass (+6-10%); the propeller grows a little (0.111-0.124 m against 0.106-0.115 m) but stays below 0.120 m in two of four arm-geometries, so with no lag reward for a small propeller the tuners still prefer one smaller than the 0.13 m published designs (post hoc: the power price and the endurance constraint, not the lag, set it).
  • Tuned for the slower lag, the designs are not on its edge, and they are robust at the faster one. At 39.1 ms, 21 of 24 still pass the margin test, with 21-48% more L2-45 crashes than at 36 ms (7l 198 and 195; lever-aware 93 and 117). Scored under the inherited 15 ms law, the 36 ms designs crash less than the designs tuned for it: 29 and 59 (7l) and 8 and 15 (lever-aware) against 106, 131, 54 and 64, at nearly the same lever-1 tracking (median J −1.70 to −1.57 against −1.81 to −1.52). Overestimating the lag cost almost nothing at 15 ms; underestimating it cost almost every flight at 36 ms.
  • Verdicts. T1 held (24 of 24). T2 held (149, 161, 63, 82 against 380, 383, 374, 376). T3 held (4 of 4). T4 held. T5 held (+14.4%, +27.5%). T6 refuted (0.121, 0.124, 0.117, 0.111 m). T7 refuted (21 of 24 pass at 39.1 ms). T8 held (kp 21.3-25.3 against 26.8-31.0). Check: 7r's designs under the inherited law reproduce its counts (106, 131, 54, 64).
  • Reading: with the motor lag taken from two published identifications instead of assumed, every design the programme had tuned since milestone 0 sits outside its stability certificate; the robustness counts of milestones 7e-7s describe vehicles with motors 2.4 times faster than the two identified 5-inch platforms. Retuning restores certified designs, and the programme's qualitative findings (lever-aware training buys robustness, robustness follows mass) hold at the sourced lag. The design lesson is asymmetric: a gain set tuned for a slower actuator than the real one loses little; one tuned for a faster actuator fails outright, so a tuner without an identified lag should assume the slow end. The programme's tuning law from here is 36 ms, D<sup>0</sup>.
  • Limits: two identifications, both from one lab's 2306-motor platforms (same motor family; brushless 5-5.1 inch only), and one includes the ESC and command path; no diameter dependence measured, so D<sup>0</sup> is a choice inside a 0.106-0.134 m range; first-order lag only (no current limit, no asymmetric rise and fall); the 36 ms midpoint is used for both vehicle and nominal model, with the plant randomisation (× 1-2) on top; six seeds, one held-out set. Simulated; preliminary.
  • Next: retrain with the lag randomised over 33-39 ms (or certified across it) and compare with tuning at the slow end, since 7t suggests the slow end alone may be enough; the 7r next steps (a flight-time requirement instead of a fixed 15 min, LiPo at about 12 min) now at 36 ms; a step-response measurement for one 4-inch and one 6-inch motor-propeller pair to give the exponent a source.

2026-11-16 · Milestone 7s: does 7r's reading survive the motor-lag law it never sourced? (predictions written and committed before experiments/m7s_lag/run.py was run on the held-out seed; outcome below)

  • Why: milestone 7r's third next step. Every design since milestone 0 flies a first-order motor-propeller lag of 0.015 (D / 0.127 m)^1.5 s, inherited and ASSUMED. 7r found both tuners shrink the propeller on the sourced closure (0.130 → 0.106-0.115 m), so the law's diameter exponent now hands the retuned designs a faster actuator (about 12 ms against 15.5 ms). If the reading rests on that, it is an artefact of an unsourced exponent.
  • Sourcing attempt: the plan was to fit the lag from listed motor constants (KV, winding resistance) and propeller masses, via the small-signal time constant τ = J / (K<sub>t</sub>²/R + 2Q/ω). The six retailer catalogues 7q read refuse connections from this session (egress proxy, 2026-11-16), and the SunnySky bench rows 7q kept carry no KV or resistance. So, as research/CLAUDE.md asks when nothing is measured, this milestone sweeps the law instead and reports whether the conclusions survive.
  • Plan (experiments/m7s_lag/run.py, no training): 7r's 72 designs (published 7l t45 cap wide and lever-aware aware cap wide under inh60; the same designs on src_li not retuned; 7r's retuned 7l src_li and aware src_li), each under the closure 7r scored it on, under nine laws: lag at 0.127 m K ∈ {7.5, 15, 30} ms × diameter exponent b ∈ {0, 1.5, 3}. The law sets both the vehicle and the controller's nominal model (aero.LAG_K, aero.LAG_EXP; default unchanged); the plant randomisation (lag × 1-2) stays on top. Held-out as 7r (seed 9999, 64 episodes), four capped conditions: lever 1 at 45°, lever 2 at 20° and 45°, lever 3 at 20°; margin test per law.
  • Looked at before writing these: a smoke run (8 episodes, seed 4242, 2 seeds per set, not the held-out seed), which showed the script runs, that at 30 ms almost no design passes the margin test, and that its few crashes rank sets differently at 30 ms than at 15 ms. It is too small to count from.
  • Predictions (L2-45 = diverged flights at lever 2, 45°, cap, summed over 6 seeds, of 384): - S0 (prior 98%): the law (15 ms, b 1.5) reproduces 7r's counts in all four conditions for all 72 designs. - S1 (prior 65%): 7i's finding survives: aware src_li crashes on fewer L2-45 flights than 7l src_li in both geometries under at least 8 of the 9 laws. - S2 (prior 60%): the mass-robustness relation survives: over the 24 published and retuned designs per geometry, Spearman ρ(mass, L2-45) < 0 in both geometries under every law with K ≤ 15 ms (6 laws). - S3 (prior 70%): slower motors crash more: at b 1.5, L2-45 rises from 7.5 to 15 to 30 ms in at least 10 of the 12 set-geometries. - S4 (prior 70%): the retuned designs lean on the exponent: going from b 1.5 to b 0 at 15 ms, 7l src_li's L2-45 rises by more than published 7l's (inh60) in both geometries. - S5 (prior 50%): under b 0 at 15 ms, 7l src_li passes the margin test in at most 8 of 12 designs (all 12 pass under the inherited law).
  • Analysis: counts per set, geometry, law and condition; margin validity; Spearman ρ per law. No test beyond counts (six seeds, one held-out set).

### Outcome (experiments/m7s_lag/results.txt, verdicts.txt; 72 designs × 9 laws × 4 conditions × 64 episodes, 888 s on 4 cores; figure m7s_lag.png; simulated)

  • 7r's two findings survive the lag law. The lever-aware retuned designs crash on fewer lever-2, 45° flights than the lever-1 retuned designs under all nine laws in both geometries (e.g. 15 ms, b 0: 85 against 133 and 92 against 143), and heavier designs crash less under every law with K ≤ 15 ms (Spearman ρ(mass, L2-45) −0.68 to −0.93 over the 24 published and retuned designs per geometry). At 30 ms the relation weakens (−0.26 to −0.59) and in one case vanishes (+0.02, above, b 3).
  • But part of 7r's parity with the published designs was the exponent. With a lag that does not depend on diameter (15 ms for every propeller), the retuned designs lose the 3-4 ms their smaller propellers bought: 7l src_li goes from 106 to 133 (above) and 131 to 143 (below) crashes, while the published 7l, on 0.13 m propellers, moves by −4 and −6. Lever-aware retuned goes from 54 and 64 to 85 and 92, against 51 and 53 for the published lever-aware designs. At equal lag, the retuned designs are less robust than the published ones, as their lower mass predicts.
  • The stability certificate is tied to the law each design was tuned on. Under the inherited law every retuned design passes the margin test (24 of 24); at b 0 only 5 of 24 do (7l 2 of 12, lever-aware 3 of 12), while the published designs pass 21 of 24. Under b 3 it is the other way round: published 7l 4 of 12, retuned 23 of 24. The tuners place each design on the stability boundary for its own lag, as milestone 7c found for the integrator gain.
  • The lag constant matters more than the exponent or the mass. Doubling K to 30 ms leaves at most 1 of 72 designs passing the margin test under each exponent and, at b 1.5, raises lever-2, 45° crashes to 244-378 of 384 in every set; halving it to 7.5 ms (b 1.5) cuts the published and retuned sets' crashes by 26-44% (the not-retuned src_li sets by 14-19%). Plotted against the lag a design actually flies, each set's crashes rise on one smooth curve (figure), so the law matters through the lag value at the design's propeller.
  • Verdicts. S0 held (0 of 288 counts differ from 7r). S1 held (9 of 9 laws). S2 held (6 of 6 laws, both geometries). S3 held (12 of 12). S4 held (+27 and +12 against −4 and −6). S5 held (2 of 12).
  • Reading: the qualitative findings (training across levers buys robustness, robustness follows mass) do not depend on the lag law over a 4× range of K and exponents 0-3. The quantitative ones do: how many flights a design loses, whether it passes the margin test, and whether a retuned design matches a published one at all depend on a constant nobody has measured for these motors. Before the programme reports counts as numbers rather than orderings, K and b need a source, and the tuner should either see the lag as uncertain (train across a lag range, as it already does across plant lag × 1-2) or be certified across it.
  • Limits: designs not retuned under each law (a tuner that knew b 0 would not shrink the propeller for lag); the law changes the controller's nominal model with the vehicle (an unidentified lag mismatch is only the plant randomisation's × 1-2); first-order small-signal lag only (no ESC current limit, no slew); the sweep range is a bracket around the inherited value, not a sourced range; six seeds, one held-out set. Simulated; preliminary.
  • Next: source K and b (motor KV and winding resistance plus propeller mass give the small-signal τ = J/(K<sub>t</sub>²/R + 2Q/ω); listings were unreachable from this session; a public step-response dataset would be better); retune with the lag randomised over its plausible range so the stability certificate covers it; the two 7r next steps (a flight-time requirement instead of a fixed 15 min; LiPo at about 12 min) still stand.

2026-11-16 · Milestone 7r: retune under the sourced, exactly solved closure, and re-ask whether training across levers buys robustness (predictions written and committed before experiments/m7r_retune/run.py was run at full budget; outcome below)

  • Why: milestone 7q's first next step. 7q found the closure the programme tuned on since milestone 0 was solved by a truncated iteration and built on constants 12-72% off listed data; with them sourced and Li-ion cells (src_li) every design closes at about 0.75 kg, and the old tunes crash 20-33 more flights of 64 per design at lever 2, 45°. Those gains were tuned on vehicles that do not exist. 7i's finding (a lever-aware objective buys lever-2 robustness by buying mass) and 7l's (a lever-1 tuner with the wide box trades robustness for a lighter airframe) both rest on that closure.
  • Plan (experiments/m7r_retune/run.py): two arms, each exactly 7l t45 cap wide or 7m aware cap wide (45° tilt limit, A_P cap, kp box 0.1-31.6 s⁻², integrator feed-forward, margin constraint, TWR 2, 3200 CMA-ES evaluations, 6 seeds × 2 starts × 2 geometries) with only the closure set to 7q's src_li (209.9 Wh/kg, motor 0.0290 (T/9 N)^1.041 kg, FM·η 0.426, SIZE_SOLVE = "exact"). A design with no fixed point sizes to the 20 kg clamp and fails the $300 cap (score −99). Held-out as 7m: 26 conditions, 64 episodes, seed 9999. The published 7l and 7m designs are re-scored under inh60 (as published) and src_li (not retuned) by the same function.
  • Looked at before writing these: a smoke run (64 evaluations, 2 seeds, 4 training episodes, scored on seed 4242, not the held-out seed) to check the script runs; its designs are unconverged and say nothing about the questions.
  • Predictions (L2-45 = diverged flights at lever 2, 45°, cap, of 64 per seed, summed over 6 seeds unless stated): - R1 (prior 95%): all 24 retuned designs are valid under src_li (close at 15 min, feasible, margin ρ < 1). - R2 (prior 60%): every arm-geometry's mean retuned mass lies in 0.65-0.90 kg. - R3 (prior 75%): in below, 7l src_li crashes on fewer L2-45 flights per design than the old 7l designs under src_li (43.0 per design over the 5 valid ones; in above none of the old ones is valid). - R4 (prior 60%): retuned, 7l src_li still crashes on more L2-45 flights than 7l under inh60 (73 and 105) in both geometries: the lighter vehicle is less robust even when tuned for. - R5 (prior 55%): 7i's question: aware src_li crashes on fewer L2-45 flights than 7l src_li in both geometries. - R6 (prior 50%): the lever-aware tuner still buys mass: aware src_li's mean mass is at least 5% above 7l src_li's in both geometries. - R7 (prior 50%): 7l src_li's median lever-1, 45° tracking J (cap) is within 0.10 of 7l under inh60 (−1.551, −1.502) in both geometries.
  • Analysis: counts per arm, geometry and condition; mass, power, arm, propeller, cant, gains; validity and emax. No test beyond counts (six seeds, one held-out set).

### Outcome (experiments/m7r_retune/results.txt, verdicts.txt; 48 CMA-ES runs of 3200 evaluations, about 2 h on 4 cores, then 72 designs × 26 held-out conditions, 652 s (the first launch hit the session's 2-hour background limit during evaluation; training was cached and only the evaluation re-ran); figure m7r_retune.png; simulated)

  • Retuned on the sourced closure, the tuners buy the mass back. All 24 retuned designs are valid (close at 15 min, feasible, margin ρ < 1), but none is the 0.75 kg vehicle the old designs sized to: 7l src_li 1.133 and 0.989 kg, aware src_li 1.289 and 1.224 kg. The lever they pull is the propeller, which they shrink from 0.130-0.134 m to 0.106-0.115 m (post hoc). A smaller disc needs more power to hover (7l 402 and 313 W against 173 and 162 W for the old designs on src_li), so the battery and then the motors grow until the vehicle again sits on the closure's edge: the longest closing endurance is 15.0-16.7 min, against 19.0-20.6 min for the old designs on the same closure.
  • Mass still tracks robustness, along one curve across closures. At lever 2, 45°, cap: the retuned designs crash on 106 and 131 (7l objective) and 54 and 64 (lever-aware) of 384, against 260 and 263 for the old 7l gains on src_li and 73 and 105 for 7l as published. Plotted against sized mass, the retuned and published designs of both objectives fall on one falling curve (figure); the not-retuned src_li designs sit off it, at 0.73-0.80 kg and 36-48 crashes of 64.
  • 7i's finding survives the sourced closure. Training across levers still buys robustness, and still by buying mass: aware src_li crashes on half 7l src_li's lever-2 flights (54 against 106; 64 against 131) and is 14% and 24% heavier. It matches the published lever-aware designs (53 and 52) at lever 2, 45°, and at lever 3, 20° (115 and 150 against 107 and 127).
  • Verdicts. R1 held (24 of 24). R2 refuted (means 0.99-1.29 kg, none in 0.65-0.90). R3 held (21.8 against 43.0 per design). R4 held (106 > 73; 131 > 105). R5 held (54 < 106; 64 < 131). R6 held (+13.7%, +23.8%). R7 held above (|ΔJ| 0.057) and refuted below (0.305).
  • Reading: 7q's "half the mass" was the old designs on the new closure, not what a tuner does with it. Asked to trade crash cost against 0.03 of the composite per 100 W of hover power, the tuner prefers a heavier, less efficient vehicle with a smaller propeller and lands on the closure's edge again, as it did on the inherited closure. So the programme's mass-robustness relation is not an artefact of the truncated solve: it reappears under the sourced, exactly solved closure, and the lever-aware objective's advantage reappears with it. What the sourced closure changes is the price: the same robustness now comes with a smaller propeller and more hover power for less mass (7l above 402 W at 1.13 kg against 340 W at 1.29 kg published; lever-aware above 495 W at 1.29 kg against 408 W at 1.45 kg).
  • Limits: the objective's power price (0.03 per 100 W) is inherited and not sourced; the endurance is a constraint, so the tuner spends any slack in it; one closure (Li-ion); the propeller-diameter range and the lag model (0.015 (D/0.127)^1.5 s) are inherited; six seeds, one held-out set. Simulated; preliminary.
  • Next: price the vehicle the way a builder does (all-up mass or hover power against a flight-time requirement rather than a fixed 15 min) and see whether the mass-robustness trade survives; retune on LiPo at the longest closing endurance (about 12 min); check the propeller-lag model against listed 4-5 inch motor-propeller step data, since the tuner now leans on small propellers.

2026-11-16 · Milestone 7q: does the sizing closure hold with component constants taken from listed data, and is it solved? (predictions written and committed before experiments/m7q_closure/run.py was run with flights; outcome below)

  • Why: milestone 7p's first next step. 7p found that the sizing closure (15 min endurance; battery 150 Wh/kg at 80% depth of discharge; motor mass 0.045 (Tmax / 9 N)^1.2 kg; propulsive efficiency FM·η = 0.65 × 0.75; all inherited and ASSUMED) multiplies a frame-mass change 6-15 times, so every mass-dependent result leans on it.
  • Data (data/components/, experiments/m7q_closure/collect.py, public Shopify catalogues accessed 2026-11-16): 405 battery listings from six FPV retailers (195 in class: 3-6S, 1000-10000 mAh, specific energy 50-320 Wh/kg; 38 implausible and 172 out of class kept but excluded, with the reason in use); 29 SunnySky motor pages carrying the manufacturer's bench tables (5,292 rows of propeller, thrust and electrical watts, each row checked by thrust / watts = the listed g/W within 5%); 16 FPV motor listings stating a maximum thrust (9 excluded by hand with the reason in fit.py: packing weights, lower bounds, a truncated sentence).
  • Fit (experiments/m7q_closure/fit.py, closure.txt): battery median 132.3 Wh/kg LiPo (n 140, CI 128.7-134.2), 156.2 LiHV (n 31), 209.9 Li-ion (n 24), against the inherited 150. Motor mass 0.0290 (Tmax / 9 N)^1.041 kg over 33 motors (bootstrap k 0.024-0.033, exponent 0.95-1.16; residual SD 0.27 in log); the inherited law is 1.72 times the listed mass at the median motor (0.90-3.08). FM·η at half each bench run's maximum thrust: 0.426 on 5-inch propellers (16 runs, 3 motors), 0.444 on 7-10 inch, 0.52-0.56 on 11-29 inch, against the inherited 0.4875.
  • Found while building the check: aero.size solves the mass closure with a damped iteration from 1 kg stopped after 60 steps. Near the closure's edge it has not converged, and where no fixed point exists it returns the mass it reached (checked on 16 designs: 60 steps against 600 and 20,000; e.g. 7e above seed 0 1.77 kg after 60 steps, the 20 kg clamp after 600). aero.SIZE_SOLVE = "exact" (new, opt-in; the default "iter60" is unchanged and reproduces 7p's 720 sizing points bit for bit) returns the smallest fixed point by scan and bisection, or the clamp when there is none.
  • Plan (experiments/m7q_closure/run.py, no training): 7o's 48 fixed designs under seven closures: inh60 (as run so far), inh (same constants, exact), bat, mot, eff (one sourced constant each, exact), src (all three, LiPo, exact), src_li (all three, Li-ion, exact). Frame law, payload, electronics, endurance, depth of discharge, thrust-to-weight and costs unchanged; designs not retuned. Per design: own-L sizing, closure gain, the longest endurance that still closes (emax), margin test, feasibility, and held-out flights (seed 9999, 64 episodes, cap on; lever 1 at 45°, lever 2 at 20° and 45°, lever 3 at 20°); 7p's arm grid (0.085-0.150 m) under inh60, inh and src_li.
  • Looked at before writing these: the sizing-only pass (results_sizing.txt), i.e. mass, power, validity and emax per design and closure. It shows that with the inherited constants solved exactly only 20 of 48 designs close at 15 min (emax medians 14.7-15.6 min: the tuned designs sit on the closure's edge), that no design closes on listed LiPo (src, emax 11.9-13.0 min), and that all close on Li-ion (src_li, 0.74-0.77 kg, emax 18.9-20.6 min) but most then fail the margin test (18 of 48 pass; with mot alone 45 of 48 pass). No crash counts were looked at.
  • Predictions (L2-45 = diverged flights at lever 2, 45°, cap, summed over the seeds valid at the point, of 64 per seed): - Q1 (prior 95%): inh60 reproduces 7o/7p at own L (7e 12 and 52, 7l 73 and 105, aware cap wide 53 and 52, fixL 39 and 46) and 7p's inh grid counts at every grid point. - Q2 (prior 70%): 7l below (all six close under inh, mass 1.102 → 1.106 kg): its inh own-L L2-45 count is within 5 of 105. - Q3 (prior 65%): under mot (about 30% lighter, not retuned), L2-45 at own L, summed over designs valid under both mot and inh60, is higher than under inh60 in at least 6 of the 8 set-geometries. - Q4 (prior 60%): under src_li, over designs valid under both, L2-45 at own L is higher than under inh60 in every set-geometry with at least two such designs. - Q5 (prior 60%): under src_li on the grid, the 7l below set's L2-45 count over seeds valid at both points does not rise by more than 2 flights between consecutive valid grid points (monotone, as 7o). - Q6 (prior 60%): median lever-1, 45° tracking J under mot is within 0.10 of inh60 for the 7l set in both geometries.
  • Analysis: counts per set, geometry, closure and L; validity; mass, power, emax; no test beyond counts (six seeds, one held-out set).

### Outcome (experiments/m7q_closure/results.txt, verdicts.txt, closure.txt; 48 designs × (7 closures at own L + 3 closures × 14 arm lengths) × 4 conditions, 980 s on 4 cores; figure m7q_closure.png; simulated)

  • The 60-step iteration hid closures that do not exist. With the inherited constants solved exactly, 28 of the 48 tuned designs have no fixed point at 15 min: no vehicle of that arm, propeller and cant carries its own battery for 15 minutes. The truncated iteration gave them 1.29-2.01 kg (median 1.40, against 1.21 for the 20 that close). All 6 heavy-basin designs (≥ 1.7 kg, the near-crash-free seeds of 7i, 7m and 7n) are among the 28. The longest endurance that closes is 14.6-15.8 min for every design: the tuners put every design on the closure's edge, where the truncated iteration's mass is largest. Where a fixed point exists, the exact mass is 0-7% above the truncated one and the L2-45 counts move by 0-3 flights per design (post hoc).
  • Listed components move every constant, and in both directions. LiPo packs store 132 Wh/kg (12% under the assumed 150); propulsion at 5 inches turns 0.426 of electrical power into ideal momentum power (13% under 0.4875); motors weigh 0.58 of the inherited law at the median motor. Either of the first two alone, or all three together on LiPo, leaves no design able to close at 15 min (emax 11.9-13.9 min). The motor fit alone closes 48 of 48 (set means 0.92-1.03 kg) (emax 15.4-16.8 min); all three on Li-ion (209.9 Wh/kg) close 48 of 48 (set means 0.74-0.77 kg) (emax 18.9-20.6 min), but only 18 of those pass the stability-margin test on their old gains.
  • On their tunes, the lighter vehicles crash far more. Paired by design, at lever 2, 45°, cap: under mot every design in every set-geometry crashes more, 4-20 more flights of 64 per design (sets: 12 → 124, 73 → 160, 53 → 124, 39 → 121 above; 52 → 107, 105 → 150, 52 → 112, 46 → 116 below); under src_li 20-33 more per design (e.g. 7l below 81 → 215 over 5 designs). Lever-1, 45° tracking J worsens by 0.59-0.80 under mot for the 7l set. The gains were tuned on heavier vehicles and are not retuned, so these counts say what the old tunes do on the sourced vehicles, not what a retuned design would do.
  • Verdicts. Q1 held (0 of 2,880 counts differ from 7p). Q2 held (106 against 105). Q3 held (8 of 8). Q4 held (5 of 5 eligible). Q5 held (largest rise 0 flights; 210 → 157 from 0.105 to 0.145 m over 5 seeds). Q6 refuted (|ΔJ| 0.80 and 0.59).
  • Reading: the robustness results since 7i rest on mass, and the mass came from a closure that (a) was not solved to convergence and (b) used component constants that, sourced, are 12-72% off. Taken together, the mission as specified (15 min on 4.8-5.3 inch propellers, 250 g payload) does not close on listed LiPo at all; it closes on Li-ion at about 0.75 kg, half the mass the tuners settled at. The heavy, robust basin is an artefact of the truncated solve. The arm's monotone effect survives the swap (Q5), so 7o's shape result holds; its level (how many crashes the arm buys back) does not carry over. The programme's next tuning runs should use SIZE_SOLVE = "exact" with the sourced closure and a mission that closes (Li-ion, or a shorter endurance on LiPo), then re-ask 7i's question.
  • Limits: listings, not measurements (battery capacities as labelled; SunnySky's bench tables are the manufacturer's; the 5-inch efficiency rests on 3 motors); the motor law is fitted to full-throttle thrust, which for FPV motors is a burst rating; the efficiency is taken at half maximum thrust (TWR 2), not at each design's own operating point; designs not retuned; six seeds, one held-out set. Simulated; preliminary.
  • Next: retune 7l and the lever-aware objective under the exact, sourced Li-ion closure (and LiPo at the longest closing endurance, about 12 min), and re-ask whether training across levers still buys robustness when mass is no longer free; make the cost of an unconverged closure explicit in every earlier milestone's README line.

2026-11-15 · Milestone 7p: does 7o's arm sweep survive a frame-mass law taken from listed frames? (predictions written and committed before experiments/m7p_frame_law/run.py was run with flights; outcome below)

  • Why: milestone 7o's third next step. 7o found the arm's long-side limit is the mass-battery spiral, and that spiral runs through the frame-mass law 0.05 + 0.8 L kg, an inherited ASSUMED constant. research/CLAUDE.md asks for driving parameters from a source, so this milestone sources it and re-runs the sweep.
  • Data (data/frames/frames_hobbyrc.csv): 39 open-propeller frame listings from one UK retailer's 3-10 inch frame categories (HobbyRC, accessed 2026-11-15), each with the verbatim wheelbase and weight line and its URL; 36 fitted, 3 excluded with the reason (a 6-inch listing repeating the 5-inch wheelbase, a 7-inch page whose weight is labelled 5-inch, a duplicate listing). Weights are as listed: some exclude TPU or printed parts (column weight_scope). Wheelbase W = 2 L.
  • Fit (experiments/m7p_frame_law/fit_frames.py, frame_law.txt): (a) one frame design listed in several sizes (GEPRC Mark4 5/6/7 with one 5 mm arm plate, Foxeer Aura V2 Pro DC 5/6): 0.255 g per mm of wheelbase, i.e. 0.51 kg per m of arm (family bootstrap 0.33-0.54), 64% of the inherited 0.8; (b) across 30 frames of 162-342 mm (20 frame lines, 7 brands): mass = -130.9 + 1.136 W g, i.e. frame = -0.131 + 2.27 L kg (bootstrap slope 1.61-2.96 kg/m; residual SD 47 g; log-log exponent 1.96), 2.8 times the inherited slope. The inherited law sits between the two: (a) is what stretching one design's arms costs, (b) what a frame designed for the larger size weighs (thicker arms, bigger plates). A wider family set (adding iFlight AOS V5 3.5/5/7, Mark4 8, Pulsar 9/10/11) gives 1.07 g/mm, dominated by AOS.
  • Plan (experiments/m7p_frame_law/run.py, no training): 7o's 48 fixed designs, full sizing (7o resize), arm L = 0.085-0.150 m in 5 mm plus own L, under four frame laws: inh (slope 0.8, = 7o), fam (slope 0.51) and xb (slope 2.27), each anchored at the design's own-L frame mass so that every design keeps the airframe it was tuned on, and xb_abs (the cross-brand law itself, which makes every design about 30 g lighter in frame at its own L; not retuned). Margin test, clearance and $300 cap per point and law. Held-out as 7o: seed 9999, 64 episodes, cap on; lever 1 at 45°, lever 2 at 20° and 45°, lever 3 at 20°.
  • Looked at before writing these: the sizing-only pass (SIZING_ONLY=1, no flights; results_sizing.txt), i.e. mass, power and validity per point and law. It shows the long-side limit is very sensitive to the slope: for 7l above, all six seeds stay valid to 0.150 m under fam (mass +71% at 0.150 m), to 0.130 m under inh (7o), and only to 0.110 m under xb (mass 1.83 kg there, runaway to 15 kg by 0.120 m). No crash counts were looked at.
  • Predictions (L2-45 = diverged flights at lever 2, 45°, cap, summed over seeds valid at the point, of 64 per seed): - P1 (prior 95%): inh reproduces 7o resize at every common grid point and at own L (7e 12 and 52, 7l 73 and 105, aware cap wide 53 and 52, fixL 39 and 46); the anchored laws give the same own-L counts. - P2 (prior 70%): under fam, every set's L2-45 count is non-increasing in L over its all-six-valid window in both geometries, allowing rises of 2 or fewer flights (no interior optimum, as 7o). - P3 (prior 50%): under fam, the 7l set's lowest all-six-valid L2-45 count is below 7o's lowest (6 above, 44 below) in at least one geometry. - P4 (prior 65%): under xb, the 7l set's lowest all-six-valid L2-45 count is above aware cap wide's own-L count (53, 52) in at least one geometry: with the cross-brand slope, 7o's finding that the arm alone matches the lever-aware co-design does not hold everywhere. - P5 (prior 55%): at grid points where a set is all-six-valid under both inh and fam, the two L2-45 counts differ by 5 or fewer flights for the 7l set in both geometries (the slope matters through validity, not through the flights). - P6 (prior 60%): xb_abs crashes more than inh at own L, L2-45, for the 7l set in both geometries (a lighter airframe on the same tune has smaller motors at fixed thrust-to-weight while the gust torque is unchanged; 7i found mass tracks robustness). - P7 (prior 60%): lever-1, 45° tracking J of the 7l set at its own L under xb_abs is within 0.05 of inh in both geometries.
  • Analysis: counts per set, geometry, law and L; validity per point; mass and power; no test beyond counts (six seeds, one held-out set).

### Outcome (experiments/m7p_frame_law/results.txt, analysis.txt, posthoc.txt; 48 designs × 15 arm lengths × 4 frame laws × 4 conditions, 742 s on 4 cores; figure m7p_frame_law.png; simulated)

  • A first run differed from 7o on 18 of 2,304 counts (by 1-2 flights, never at a design's own L). The cause was the script: inh was anchored as 0.05 + 0.8 L0 − 0.8 L0, which is not bit-exact in floating point, and 7o showed a sub-micron change flips marginal flights. inh now sets the constants exactly; the rerun differs on 0 of 2,304. Only the rerun is kept.
  • The frame slope decides where the arm stops, more than anything 7o measured. The longest arm at which all six seeds stay valid, inh → fam → xb: 7l above 130 → 150 → 110 mm, below 150 → 150 → 115; 7e above 105 → 120 → none; aware cap wide 125/130 → 135/140 → none. With the stretched-design slope the mass-battery spiral 7o named as the long-side limit mostly moves off the grid; with the cross-brand slope it arrives 15-35 mm sooner.
  • With the sourced stretch slope, a longer arm buys much less robustness. 7l at L2-45, inh against fam at the same arm: above 115 mm 31 against 48, 125 mm 14 against 30, 130 mm 6 against 21; below 135 mm 44 against 66, 150 mm 27 against 54. The curves still fall monotonically (no rise after a minimum in any of 8 set-geometries). fam's lowest counts are 5 (above, 145 mm, 2.03 kg) and 54 (below, 150 mm, 1.36 kg), against 7o's 6 and 27.
  • Part of 7o's "arm" effect was mass (post hoc). The anchored laws give the same design at the same arm different masses. At the same arm, the heavier law (inh) crashes on fewer flights than fam at 253 of 350 valid design-points above the own arm, equal at 92, more at 5 (median mass difference 95-165 g). A per-design fixed-effects fit over all valid points of the three anchored laws gives −0.14 to −0.18 flights (of 64) per mm of arm and −0.15 to −0.89 per 100 g of mass, both intervals excluding zero in all 8 set-geometries (7e above: −0.06 and −0.03). On 7l below, 105 → 135 mm under inh adds 268 g; at these rates the mass accounts for roughly a quarter of the 59-flight drop (about 14 flights).
  • The sizing closure amplifies the frame constant 6-15 times. Under the cross-brand law itself (xb_abs), each design's frame is 20-43 g lighter at its own L, and the sized vehicle 0.18-0.64 kg lighter (7l 1.29 → 1.00 and 1.10 → 0.92 kg; hover power 340 → 229 and 259 → 198 W): the battery and motor masses follow the frame through the endurance closure. On the same tunes, not retuned, the lighter vehicles crash on 160 and 151 of 384 at L2-45 (against 73 and 105) and track worse at lever 1, 45° (J −2.26 and −2.16 against −1.58 and −1.71).
  • Verdicts. P1 held (after the fix; 0 of 2,304 differ, own-L counts identical under the anchored laws). P2 held (8 of 8). P3 held, above only (5 < 6; below 54 > 44). P4 refuted (xb 7l best 15 and 41, under 53 and 52: the steeper slope adds mass faster, which buys robustness faster while the design is valid). P5 refuted (differences up to 27 flights). P6 held (160 > 73; 151 > 105). P7 refuted (J off by 0.45-0.68, far more than 0.05).
  • Reading: 7o's two conclusions about the arm both lean on the inherited frame slope. Where the arm stops is set by the mass-battery spiral, and the spiral's onset moves 40 mm across the range of sourced slopes (one design stretched against a frame designed for the size). How much robustness the arm buys depends on how much mass it brings: with the stretched-design slope, which is the honest one for lengthening one frame's arms, the gain per millimetre is smaller and 7o's "arm alone matches the lever-aware designs" holds above (48 ≤ 53 from 115 mm) but fails below by two flights at the end of the grid (54 against 52 at 150 mm). Robustness on a fixed tune tracks both arm and mass, consistent with 7i's finding that mass tracks robustness. The larger finding is about the model: the inherited endurance closure (15 min, 150 Wh/kg, the motor-mass fit) multiplies any frame-mass change 6-15 times, so every mass-dependent result in this programme carries that sensitivity until the closure itself is checked against built vehicles.
  • Limits: one retailer's listings (weights as listed, some excluding TPU or printed parts, residual SD 47 g across brands); the stretched-design slope rests on two families (5 frames); the frame law is linear in L; the designs are not retuned under the new laws; six seeds, one held-out set. Simulated; preliminary.
  • Next: check the sizing closure against built quadrotors (all-up weight, battery and motor masses of listed 5-7 inch builds; the battery's specific energy and the motor-mass fit are inherited); retrain under the fam law to see whether the tuner still lands where 7l and 7m did; a crash-weighted objective with the arm free and a hover-power budget (7o's first next step).

2026-11-15 · Milestone 7o: crashes against frame arm length on fixed tunes, and what the extra gust exposure costs (predictions written and committed before experiments/m7o_arm_sweep/run.py was run at full budget; outcome below)

  • Why: milestone 7n's first next step. 7n measured one arm swap per design (7l's designs lengthened by 4-7 mm: lever-2 crashes at 45° with the cap 73 → 59 and 105 → 87) and found that shortening 7m's arms pushed half of them past the stability margin. A sweep maps the whole curve on fixed tunes: where crashes bottom out, where the margin test and the feasibility rules (propeller clearance √2 L ≥ 1.05 D, $300 cost cap) bound it, and how much of the arm's benefit the sizing model takes back through the gust exposure that grows with L (frontal area 0.02 + 0.5 L², centre-of-pressure lever 0.05 + 0.3 L, both ASSUMED and inherited).
  • Plan (experiments/m7o_arm_sweep/run.py, no training): every design of 7e, 7l t45 cap wide, 7m aware cap wide and 7n aware cap wide fixL (6 seeds × 2 geometries each, 48 designs) with only its arm set to L = 0.085-0.135 m in 5 mm steps, plus its own L. Two sizing modes: resize (the full model: mass, inertia, torque limit, hover power, area and lever all follow L) and pinexp (area and lever held at the design's own-L values, so only the actuator side moves). Each point gets the margin test under its own kp box and the feasibility rules; only points passing both are counted. Held-out, 7m's seed 9999 and 64 episodes, with the cap: lever 1 at 45° (tracking), lever 2 at 20° and 45°, lever 3 at 20°.
  • Looked at before writing these: one smoke run (2 seeds, 8 episodes, seed 4242, not the held-out seed; 50 s). It showed that on the heavy 7e designs the sizing closure runs away as L grows (mass 1.8 kg at 0.100 m to the 20 kg clamp by 0.130 m in above, the battery feeding back on the frame mass), which the cost cap then marks infeasible; that is why feasibility, not only clearance, gates the counts. Its counts are too small to say anything about the questions.
  • Predictions (L2-45 = diverged flights at lever 2, 45°, cap, summed over the seeds that are ok at the point, of 64 per seed): - O1 (prior 90%): at each design's own L the resize counts reproduce 7m and 7n at L2-45: 7e 12 and 52, 7l 73 and 105, aware cap wide 53 and 52, fixL 39 and 46. - O2 (prior 65%): resize, 7l set: L2-45 at 0.115 m is below L2-45 at 0.100 m in both geometries (over seeds ok at both). - O3 (prior 50%): resize: in at least one set and geometry the L2-45 curve over the ok grid has an interior minimum, i.e. crashes rise again before feasibility ends the sweep. - O4 (prior 60%): pinexp crashes fewer flights than resize at L2-45 summed over the grid points above each design's own L, for the 7l set in both geometries (the extra gust exposure takes back part of the gain). - O5 (prior 70%): at 0.085 m at least half of all 48 designs fail the margin test or a feasibility rule; at every grid point at or above a design's own L, every design passes the margin test. - O6 (prior 45%): at its best ok grid point (all six seeds ok), the 7l set reaches an L2-45 count at or below 7m aware cap wide's own-L count (53 and 52) in both geometries: on these tunes the arm alone can close the gap the objective opened. - O7 (prior 60%): resize lever-1 45° tracking J for the 7l set stays within 0.05 of its own-L value over 0.100-0.120 m in both geometries.
  • Analysis: counts per set, geometry, mode, condition and L; ok counts; mass, torque per weight and exposure along the sweep; per-seed curves in results.json. No test beyond counts: six seeds, one held-out set.

### Outcome (experiments/m7o_arm_sweep/results.txt and analysis.txt, 48 designs × 12 arm lengths × 2 sizing modes × 4 conditions, 347 s on 4 cores; figure m7o_arm_sweep.png; simulated)

  • A first run had own-L counts off by 1-2 (e.g. 7l below 106 against 105). The cause was the script, not the simulator: it rounded each design's own L to 6 decimals before re-encoding it, and a sub-micron change in L flipped one marginal flight (checked: same design, coordinate untouched 14, rounded 15). The script now scores each design's own coordinate untouched and was rerun; grid points were unaffected. Worth knowing for any later swap study: marginal flights flip on tiny design changes, so single-flight differences mean nothing.
  • Baselines reproduced exactly. At L2-45 (own L): 7e 12 and 52, 7l 73 and 105, aware cap wide 53 and 52, fixL 39 and 46.
  • Crashes fall with the arm across the whole valid range, with no interior optimum. Over the window where all six seeds of a set are valid, every set's L2-45 count is non-increasing in L (7l above 62, 52, 31, 21, 14, 6 from 0.105 to 0.130 m; below 103, 87, 74, 70, 61, 54, 44 from 0.105 to 0.135 m; 7e below 55 → 5 from 0.100 to 0.135 m). Per design, none of 46 with at least three valid points has a curve that rises by 3 or more flights after its minimum, in either sizing mode.
  • What bounds the arm is validity, not robustness. Short side: at 0.085 m, 47 of 48 designs fail (34 the margin test, 45 propeller clearance or cost), and the margin test fails only below a design's own L (0 failures at or above it in 48 designs). Long side: the sizing closure. The frame's 0.8 kg/m feeds the battery and motor masses, so mass rises faster than linearly: 7l above +3% at 0.105 m, +26% at 0.120, +98% at 0.135 (hover power +5%, +42%, +188%); 7e above runs away to the 20 kg clamp by 0.130 m and hits the $300 cap from 0.110-0.120 m. 49 grid points at or above a design's own L are infeasible for cost, 29 of them 7e above and 15 the heavy-basin fixL designs.
  • The added gust exposure takes back part of the gain. With area and centre-of-pressure lever held at their own-L values (pinexp), the 7l set crashes on 123 and 324 flights summed over its valid points above own L, against 190 and 493 with the full model. At the longest all-valid arm: 7l above (0.130 m) 6 against 1, below (0.135 m) 44 against 12; aware cap wide 9 against 2 and 14 against 2. In below the exposure takes back about a third of the drop from own L (105 → 44 full, 105 → 12 pinned).
  • On these tunes the arm alone matches the lever-aware co-design. 7l's designs reach aware cap wide's own-L count at 0.110 m in above (52 against 53; mean 1.411 kg and 388 W against 1.446 kg and 408 W) and at 0.135 m in below (44 against 52; 1.384 kg and 367 W against 1.397 kg and 387 W), with the same median lever-1, 45° tracking J (-1.548 against -1.543; -1.521 against -1.514) and the same lever-1 crashes (0 and 0; 1 and 1). Lengthening 7l's arm to 0.135 m is a large geometric change (33%) that the tuners never chose.
  • Tracking cost is not flat. 7l's lever-1, 45° J moves by -0.038 to +0.205 (above) and -0.547 to +0.688 (below) from its own-L value over 0.100-0.120 m: single crashed lever-1 flights swing a seed's mean, and the hover-power term grows with the mass.
  • Verdicts. O1 held (after the rounding fix). O2 held (33 → 9 and 52 → 37 over the 3 seeds valid at both points in each geometry; fewer in all 6). O3 refuted (no interior minimum anywhere). O4 held (123 < 190; 324 < 493). O5 held (47 of 48 fail at 0.085 m; no margin failure at or above own L). O6 held (6 ≤ 53; 44 ≤ 52), the below match only at the end of the grid. O7 refuted.
  • Reading: on this sizing model a longer arm is a monotone robustness knob on a fixed tune, and what limits it is the design staying valid: the stability margin below the tuned length and the mass-battery spiral above it. Its gust-exposure side effect is real (a third of the benefit in below) but never outweighs the torque it buys. 7n showed the lever-aware tuner does not need the arm; 7o shows the arm alone, applied to the lever-1 tune, can reach the lever-aware designs' robustness at about the same mass and power. The tuners stop where they do because their training score barely sees lever 2 (7l) or prices mass through hover power (both), not because longer arms stop helping. For a builder: lengthen arms on an existing tune rather than shortening them (shorter fails the margin test), and stop where hover power starts to climb steeply (about +40% by 0.120 m in above).
  • Limits: frontal area and CoP lever laws are inherited assumptions (pinexp brackets their effect; the true dependence is likely between the two modes); the frame mass is linear in L; one held-out set, six seeds, and set sums over designs that are valid at different L; the arm change is evaluated, not retrained (a retrain at each L would re-tune the gains). Simulated; preliminary.
  • Next: a crash-weighted or lever-2 objective with the arm free and a hover-power budget, to see whether the tuner then chooses the long arm; retrain at fixed arms 0.110-0.130 m to separate the arm from the tune; a sourced frame-mass and frontal-area law (commercial 5-7 inch frames) in place of the inherited linear one.

2026-11-15 · Milestone 7n: is the longer frame arm what buys 7m's lever-2 robustness? (predictions written and committed before experiments/m7n_fixed_arm/run.py was run at full budget; outcome below)

  • Why: milestone 7m's first next step. 7m's aware cap wide designs (lever-aware objective, wide position-kp box, cap) crashed on fewer lever-2 flights than 7l's t45 cap wide (53 and 52 against 73 and 105 of 384 at 45° with the cap) and had a longer frame arm (mean L 0.106 and 0.110 m against 0.102 and 0.103). 7m read the arm as the free way to raise torque per weight, but from the chosen designs, not a controlled run. Two things in that reading need checking: seed 3's heavy basin lengthens the arm most (0.116 and 0.122 m) and so carries much of the mean, and on this sizing model the arm also lengthens the centre-of-pressure lever (aero.size: lever = 0.05 + 0.3 L), so a longer arm raises the gust torque as well as the control torque.
  • Plan (experiments/m7n_fixed_arm/run.py): (1) retrain: 7m's aware cap wide exactly (lever-aware score at levers 0, 1, 2, 6 episodes each; 45° tilt limit; cap A_P = 3.5 m/s²; kp box 0.1-31.6 s⁻²; 3200 CMA-ES evaluations, 6 seeds, 2 starts, both geometries), with the arm pinned at 7l t45 cap wide's mean L for the geometry (0.1017 and 0.1027 m). CMA-ES still carries the pinned coordinate; it is overwritten before scoring. The propeller-diameter start moves from 0.5 to 0.25 of its range (D 0.165 to 0.1225 m), because 7m's start is infeasible at the pinned arm (√2 L ≥ 1.05 D); about 95% of the first candidates fail the margin test at either start. (2) Swap, no training: each 7m aware cap wide design with only its arm set to the pin (7m short arm), and each 7l t45 cap wide design with only its arm set to 7m's mean (0.1062 and 0.1096 m; 7l long arm). Each design's margin test (margin_rho under its own kp box) is reported; a swapped design that fails it could not have been picked by the tuner. Held-out: 7m's 26 conditions, 64 episodes, seed 9999.
  • Looked at before writing these: two smoke runs (64 evaluations, 2 seeds, 2 episodes per lever, scored on seed 4242, not the held-out seed). They found the infeasible start and a margin-test bug in the script (tested under the default kp box), both fixed; their designs are unconverged and say nothing about the questions.
  • Predictions (counts are diverged held-out flights of 384 per set and geometry; "cap" = evaluated with the cap; L2-45 = lever 2, 45°, cap): - N1 (prior 90%): the re-scored sets reproduce 7m at L2-45: 7e 12 and 52, 7l t45 cap wide 73 and 105, aware cap wide 53 and 52. - N2 (prior 60%): aware cap wide fixL crashes fewer flights than 7l t45 cap wide at L2-45 in both geometries (pinning the arm does not undo what the lever-aware objective buys). - N3 (prior 55%): aware cap wide fixL crashes no more than 1.25 × aware cap wide's count at L2-45 in both geometries (66 and 65). - N4 (prior 60%): aware cap wide fixL is heavier than aware cap wide (1.446 and 1.397 kg) in at least one geometry (with the arm fixed, the tuner buys torque with mass). - N5 (prior 65%): the 7m short arm swap moves the L2-45 count by less than a quarter of the 7m-7l gap in both geometries (fewer than 5 more crashes in above, fewer than 14 in below). - N6 (prior 55%): the 7l long arm swap lowers the L2-45 count in both geometries, but by less than half the 7m-7l gap. - N7 (prior 70%): aware cap wide fixL's mean position kp is above 15 s⁻² in both geometries. - N8 (prior 60%): at least one swapped design fails the margin test (every design sits near rho = 1, and the arm moves the inertia).
  • Analysis: counts per set, geometry and condition; paired dC per seed against 7e; for the swaps, the per-seed change in L2-45 crashes against the unswapped design; mass, torque per weight, gains, margin rho. No test beyond counts: six seeds, one held-out set.

### Outcome (experiments/m7n_fixed_arm/results.txt, 24 CMA-ES runs of 3200 evaluations and 72 designs × 26 held-out conditions, 4,155 s on 4 cores; figure m7n_fixed_arm.png; simulated)

  • Baselines reproduced. At L2-45: 7e 12 and 52, 7l t45 cap wide 73 and 105, aware cap wide 53 and 52.
  • Pinning the arm costs no robustness. aware cap wide fixL (arm 0.102 and 0.103 m) crashes on 39 and 46 of 384 at L2-45, against 53 and 52 with the free arm and 73 and 105 for 7l. Per seed against aware cap wide: fewer in 4 of 6 and equal in 2 (above); fewer in 3, equal in 2, more in 1 (below; seed 5, 8 → 9); 9 of above's 14-crash difference is seed 5. Lever 2, 20°, cap: 19 and 33 (free arm 21 and 32); lever 3, 20°, cap: 98 and 129 (107 and 127). Position kp 21.9 and 22.4 s⁻² (free arm 23.8 and 24.3); mass 1.496 and 1.397 kg (1.446 and 1.397); torque per newton of weight 0.0682 m in both (0.0702 and 0.0728).
  • What it does instead. Seed 3 again lands in the heavy basin (1.99 and 2.01 kg, 0 crashes at L2-45), and above seed 5 finds a second heavy design (1.75 kg, kp 10.9 s⁻², propeller 93.5 mm, cant 0.3°, i.e. almost uncanted) with 0 crashes and poor tracking (held-out C 3.21 at lever 1, 20°, cap, against 1.42-1.47 for the other five), which the three-lever training score accepted. Without those two the above sets are close: 39 against 44 crashes over seeds 0-2 and 4 (free arm: 13, 13, 10, 8 against pinned 10, 12, 9, 8).
  • On the same design, a longer arm helps a little, and it is not free. 7l long arm (each 7l design with only L raised to 0.106 and 0.110 m): L2-45 73 → 59 and 105 → 87, fewer crashes in 11 of 12 seeds (equal in 1); lever 3, 20°, cap 142 → 111 and 223 → 200; tracking at lever 1, 45°, cap no worse (dC -0.211 to +0.001, lower in 11 of 12). On this sizing model the arm is not mass-free: the frame mass grows with L, so the swap adds 0.058 and 0.044 kg (4%) and 7% and 6% hover power, while torque per weight rises 0.067 → 0.070 and 0.068 → 0.073 m. 7m's line that a longer arm "raises torque without mass" was wrong; it raises torque faster than mass.
  • Shortening 7m's arms mostly makes designs the tuner could not have picked. Of the 12 7m short arm designs, 6 fail the margin test (rho 1.0001-1.0008; all 6 long-arm swaps of 7l pass), and above seed 0 also breaks the propeller-clearance rule (√2 L < 1.05 D: D 139 mm at L 102 mm), so its J is the infeasible -99. Over all 12: 53 → 67 and 52 → 65 at L2-45, with seed 3's heavy design losing 0.49 and 0.60 kg along with the arm (0 → 8 crashes in each geometry). Over the 6 designs that pass both tests: per-design changes -2 to +1 (above seeds 1, 2: 13 → 11, 10 → 10; below seeds 0, 1, 2, 5: +1 each).
  • Verdicts. N1 held. N2 held (39 < 73; 46 < 105). N3 held (39 and 46, below the free arm's own 53 and 52). N4 held in above only (1.496 > 1.446; below equal at 1.397), and through seed 5. N5 refuted over all designs (above +14 against a 5 threshold; below +13 under 14); over the 6 valid swaps it would hold. N6 refuted (above fell 14, more than half the 20 gap; below fell 18, under half of 53). N7 held (21.9, 22.4). N8 held (6 short-arm swaps fail the margin; one also the clearance rule, which was not predicted).
  • Reading: 7m's arm is a symptom, not the cause. With the arm held at 7l's length, the lever-aware tuner reaches the same or better lever-2 robustness through mass and gains (two heavy basins, slightly lower kp), so what buys 7m's robustness over 7l is the objective, not the arm. Locally the arm does work in the expected direction: lengthening a design's arm by 4-7 mm removes about one crash in five at twice the lever for about 4% more mass, a cheaper trade per kilogram than the mass route 7i found. For a builder: a few millimetres of extra arm is a cheap robustness margin on a fixed tune, but it is not the reason a robustly-trained co-design is robust, and shortening an arm on a tune sitting at its stability margin can push it past that margin.
  • Limits: the pin is one value per geometry (7l's mean), not a sweep; seed-level basins (seeds 3 and 5) carry the set means; the swap changes mass with the arm, so it is not an arm-only change in the vehicle, only in the design vector; six seeds, one simulator, 64 held-out flights per condition. Simulated; preliminary.
  • Next: an arm sweep on fixed tunes (L from 0.09 to 0.13 m, re-sizing mass) to map crashes against torque per weight directly; a crash-weighted objective that does not accept seed 5's poor-tracking design; sweep the cap (2-6 m/s²) under the lever-aware objective; start the tuner from the heavy basin on purpose.

2026-11-15 · Milestone 7m: the wide position-gain box and the proportional-request cap under the lever-aware objective: does the stiffness still buy a lighter airframe? (predictions written and committed before experiments/m7m_aware_wide/run.py was run at full budget; outcome below)

  • Why: milestone 7l's first next step. 7l found that with |kp e| capped at 3.5 m/s² and the position-kp box widened to 31.6 s⁻², a lever-1 tuner raises kp to about 27 s⁻² and converts it into a 19% (above) and 11% (below) lighter airframe, which then crashes more at lever 2 even with the cap (73 and 105 of 384 at 45° against 7e's 12 and 52). 7i found the opposite pull: a lever-aware objective (levers 0-2 in the training score) makes the tuner buy mass, which on this sizing model is torque authority against a mass-independent gust torque (below 1.23 -> 1.49 kg, lever-2 crashes 45 -> 20 at 20°). Which wins when both are present?
  • Plan (experiments/m7m_aware_wide/run.py): 7i's lever-aware training score (mean of poshold.score at levers 0, 1, 2, 6 episodes each) with 7l's cap arms' setting (tilt-command limit 45°, cap A_P = 3.5 m/s² as a per-design position-error clip A_P / kp). Arms: aware cap (kp box 0.1-10 s⁻²) and aware cap wide (0.1-31.6 s⁻²). Otherwise 7e's training (integrator feed-forward, ki <= ceiling / 1.25 as a constraint, thrust-to-weight 2, 3200 CMA-ES evaluations, 6 seeds, 2 starts, both rigid geometries). Baselines re-scored by the same function: 7e, 7i aware TWR=2.0 (7i aware), 7l t45 cap wide. Held-out: 7l's 22 conditions plus lever 3 (outside all training) at 20° and 45°, without and with the cap (64 episodes, seed 9999). New in the summary: torque limit and torque per newton of weight, frame arm and propeller diameter.
  • Looked at before writing these: one smoke run (64 evaluations, 2 seeds, 2 episodes per lever, scored on seed 4242, not the held-out seed; 230 s). Its designs are unconverged (0.8 kg, kp 0.5-1.9) and say nothing about the questions; it checked the pipeline end to end.
  • Predictions (counts are diverged held-out flights of 384 per arm and geometry; "with the cap" = evaluated with the cap): - M1 (prior 90%): the three baselines reproduce their milestones' counts (7e lever 2, 20°: 4 and 45; 7i aware 7 and 20; 7l t45 cap wide lever 2, 45°, cap: 73 and 105). - M2 (prior 70%): aware cap wide picks a mean position kp above 15 s⁻² in both geometries (the lever-aware objective does not undo 7l's stiffness). - M3 (prior 60%): aware cap wide is heavier than 7l t45 cap wide in both geometries (1.29 and 1.10 kg). - M4 (prior 45%): aware cap wide is at least as heavy as 7e in both geometries (1.61 and 1.23 kg): the lever-aware pull wins outright. - M5 (prior 65%): at lever 2, 45°, with the cap, aware cap wide crashes on fewer flights than 7l t45 cap wide in both geometries. - M6 (prior 50%): at lever 2, 20°, with the cap, aware cap wide crashes on no more flights than 7i aware in below (20), i.e. the extra stiffness costs nothing at the lever it was trained on. - M7 (prior 55%): aware cap (inherited box) keeps kp at 9.9-10 s⁻² in every seed of both geometries, as every earlier arm did.
  • Analysis: counts per arm, geometry and condition; paired dC per seed against 7e; mass, hover power, torque limit, torque per weight and the chosen gains. No test beyond counts: six seeds, one held-out set.

### Outcome (experiments/m7m_aware_wide/results.txt, 48 CMA-ES runs of 3200 evaluations and 60 designs × 26 held-out conditions; about 2 h 10 min on 4 cores over two launches: the first hit the session's 2-hour background limit after 32 of 48 runs, which were cached, and the resumed run reports its own 4,073 s; figure m7m_aware_wide.png; simulated)

  • Baselines reproduced. 7e at lever 2, 20°: 4 and 45; 7i aware: 7 and 20; 7l t45 cap wide at lever 2, 45°, cap: 73 and 105.
  • The stiffness survives the lever-aware objective. aware cap wide position kp 19.9-27.3 s⁻² (above, mean 23.8) and 19.8-31.2 (below, mean 24.3), against 7l's 27.5 and 27.2. aware cap (old box): 9.95-10.00 in all 12 seeds.
  • Mass splits the difference; the arm lengthens. aware cap wide 1.446 kg (above) and 1.397 kg (below), hover 408 and 387 W; 7l t45 cap wide 1.294 and 1.102 kg; aware cap 1.599 and 1.526 kg; 7e 1.605 and 1.234 kg. Frame arm L 0.106 and 0.110 m, against 0.094-0.103 m for every other set; torque limit per newton of weight 0.0702 and 0.0728 m against 0.0622-0.0679. At fixed thrust-to-weight a longer arm is the one sizing handle that raises torque without mass, and only this arm reaches for it.
  • Best tracking at the trained levers. Lever 1, 20°, cap: C minus 7e -0.030 (se 0.010, better in 5 of 6) and -0.044 (se 0.010, 6 of 6); at 45°, cap: -0.034 and -0.048. No other set gains more than 0.013.
  • Robustness in between. Lever 2, 45°, cap: 53 and 52 (7l 73 and 105; aware cap 19 and 26; 7i 16 and 24; 7e 12 and 52). Lever 2, 20°, cap: 21 and 32 (7i 6 and 20; aware cap 11 and 18). Lever 3, 20°, cap: 107 and 127 (7l 142 and 223; 7e 69 and 158; 7i 66 and 87).
  • Still cap-critical. No cap at 45°: 234 and 231 crashes at lever 1, 250 and 239 at lever 2 (7l 267-284). With the cap at a 1.0 m offset: 0 and 0.
  • Seed 3 again. As in 7i, seed 3 of every lever-aware arm lands one heavy design (1.94-2.02 kg) with 0 or 1 of 64 crashes at lever 2 with the cap (the aware cap wide ones still crash on 25-47 at 45° without it); it sets much of each lever-aware arm's mean. The other five aware cap wide seeds weigh 1.15-1.43 kg.
  • Verdicts. M1 held. M2 held (23.8, 24.3). M3 held (1.446 > 1.294; 1.397 > 1.102). M4 refuted (above 1.446 < 1.605; below held, 1.397 > 1.234). M5 held (53 < 73; 52 < 105). M6 refuted (32 against 20). M7 held (9.95-10.00).
  • Reading: when the objective asks for robustness to the centre of pressure, the co-design still takes most of the stiffness a wider box offers, but it no longer spends all of it on mass: it keeps part of the mass and lengthens the frame arm, the only free way on this sizing model to raise torque per unit weight. It also tracks best of any set at the trained levers. The result is a design between 7l's and 7i's: lever-2 crashes fall by a quarter (above) to a half (below) from 7l's, but stay above the lever-aware designs at the old box. Without the cap it is as fragile as 7l's. For a builder the cap stays safety-critical whatever the objective. The lever-aware objective decides how much of the gain-box room becomes a lighter vehicle.
  • Limits: the arm-length reading is from the chosen designs, not a controlled run with the arm fixed; seed 3's basin dominates the means; the cap value is 7k's and not swept; six seeds, 6 training episodes per lever, one simulator, 64 held-out flights per condition. Simulated; preliminary.
  • Next: a fixed-arm control run to test whether the arm is what buys the robustness; sweep the cap (2-6 m/s²) with the wide box under this objective; start the lever-aware tuner from the seed-3 basin; a crash-weighted objective.

2026-11-15 · Milestone 7l: retrain at the tilt limit you fly, with and without the proportional-request cap (predictions written and committed before experiments/m7l_train45/run.py was run at full budget; outcome below)

  • Why: milestones 7j and 7k's first next step. 7j advised "tune at the limit you fly" for the 45° opening-transient crash; 7k found a controller-side fix (bound the proportional position request at the lean the tune saw) and noted that every design sits at the 10 s⁻² top of the position-kp box. Does training at 45° remove the crash by itself, and at what price? With the cap in the loop during training, does the tuner keep (or, with a wider box, raise) kp, since the cap removes the transient a high kp causes?
  • Plan (experiments/m7l_train45/run.py): milestone 7e's training exactly (lever-1 objective, 16 episodes, integrator feed-forward, ki <= ceiling / 1.25 as a constraint, thrust-to-weight 2, 3200 CMA-ES evaluations, 6 seeds, 2 starts, both rigid geometries) with the tilt-command limit at 45° (PX4's MPC_TILTMAX_AIR default). Arms: t45 (no cap), t45 cap (|kp e| <= A_P = 3.5 m/s², 7k's 0.35 m × 10 s⁻², about the 20° lean g tan 20° = 3.57 m/s², applied as a per-design position-error clip A_P / kp so the bound is an acceleration whatever kp the tuner picks), t45 cap wide (the same cap, kp box widened from 0.1-10 to 0.1-31.6 s⁻²; the CMA-ES start then maps to kp 1.78 instead of 1.0). Baseline: 7e's constrained designs (trained at 20°), re-scored by the same function. Held-out: 64 episodes (seed 9999), levers 0-2, tilt limits 20°, 30°, 45°, each without and with the A_P cap at evaluation; 45° at a 1.0 m start offset, levers 0 and 1. poshold.score gains a perr_max passthrough, inert at None and tested with a per-design clip (tests/test_poshold.py).
  • Looked at before writing these: one smoke run (64 evaluations, 2 seeds, 4 training episodes, scored on seed 4242, not the held-out seed; 117 s). Its designs are far from converged (held-out C 6-12 against 7e's 1.5) and say nothing about the questions; it checked the pipeline and timed it (full run estimated near 100 minutes on 4 cores).
  • Predictions (counts are diverged held-out flights out of 384 per arm and geometry, 0.5 m offset unless stated): - L1 (prior 90%): 7e reproduces the earlier counts: lever 0, 45°, no cap: above 0, below 1; lever 2, 20°: 4 and 45; lever 2, 45°: 12 and 55. - L2 (prior 65%): t45 crashes on 0 flights at 45° at levers 0 and 1 in both geometries (tuning at the limit removes the transient crash by itself). - L3 (prior 55%): t45's mean position kp is below 7e's in both geometries (without the cap, the tuner pays for the transient by backing off kp). - L4 (prior 55%): t45 cap's mean position kp is at least 9 s⁻² in both geometries (with the cap, high kp costs nothing in the transient). - L5 (prior 45%): t45 cap wide's mean position kp is above 12 s⁻² in both geometries (off the old box edge). - L6 (prior 55%): at 45°, lever 1, t45 cap wide (evaluated with its cap) has a lower mean held-out C than t45 (no cap) in both geometries. - L7 (prior 60%): at a 1.0 m offset, 45°, levers 0 and 1, both cap arms evaluated with their cap crash 0 flights in both geometries, and t45 without a cap crashes at least 1 in at least one geometry. - L8 (prior 55%): no 45°-trained arm cuts the below lever-2, 45° count to less than half of 7e's (7g: those tip-overs are torque authority, not the tune). - L9 (prior 60%): flown at 20° (lever 1, no cap), every 45°-trained arm costs more than 7e (mean paired dC > 0) in both geometries.
  • Analysis: counts per arm, geometry and condition; paired dC per seed against 7e (mean, se, how many worse); the chosen gains. No test beyond counts: six seeds, one held-out set.

### Outcome (experiments/m7l_train45/results.txt, 72 CMA-ES runs plus 48 designs × 22 held-out conditions, 101 min on 4 cores; figure m7l_train45.png; simulated)

  • 7e reproduced. Lever 0, 45°, no cap: above 0, below 1; lever 2, 20°: 4 and 45; lever 2, 45°: 12 and 55.
  • Training at 45° buys nothing over the 20° tune. t45 keeps every gain except a larger position ki (1.3-1.8 s⁻³ in 5 of 6 above seeds against 0.01-0.55 for 7e) and lands on 7e's cost: paired dC against 7e within ±0.005 at 20°, 30° and 45° (levers 0 and 1, above), and at most -0.32 in below, where 7e's 6 crashes at 45°, lever 1 become 3. Like 7e, it is not the design set that crashed in 7j. At a 1.0 m start offset it crashes as often as 7e (92-103 against 97-118 of 384): tuning at the limit does not teach a larger reposition. The cap does, for either set (0 of 384).
  • The tuner does not back off position kp, with or without the cap. t45 and t45 cap both sit at 9.96-10.00 s⁻², as 7e does. Trained with the cap in the loop, t45 cap is no better than 7e with the cap added afterwards (45°, lever 1, capped: 0 and 1 crashes, C 1.463 and 1.545 against 7e's 1.464 and 1.553) and, flown without it, worse than 7e (above 3-6 crashes at 45° against 0; 123-129 at 1.0 m against 97-101). The tune comes to lean on the cap.
  • Given room, the tuner raises kp to the new edge and spends it on a lighter vehicle. t45 cap wide picks kp 21.7-31.5 s⁻² (mean 27.5 above, 27.2 below; two seeds within 0.5 of the 31.6 edge). Its designs are 19% lighter in above (1.29 against 1.61 kg) and 11% in below (1.10 against 1.23 kg), with 30% and 18% less hover power (340 against 486 W; 259 against 316 W). At the trained lever with the cap they hold position as well as 7e (C at 20°, lever 1: 1.455 and 1.449 against 1.467 and 1.451) and score better J for the power saved (-1.557 and -1.527 against -1.613 and -1.546).
  • The price is robustness, and the cap becomes safety-critical. Without the cap at 45°, the wide designs crash on 267-274 of 384 flights in both geometries (70%), and on 32-62 at 30°. With it they still crash more than 7e at twice the lever: 45°, lever 2, capped: 73 and 105 against 12 and 52; at 20° without a cap (where the per-axis clip bites): 35 and 69 against 4 and 45. At a 1.0 m offset, capped: 0-6 crashes against 7e's 0.
  • Verdicts. L1 held. L2 refuted (below: 3 crashes at 45°, lever 1). L3 refuted (kp 9.98-10.00 in t45, as 7e). L4 held. L5 held (27.5 and 27.2). L6 refuted (above: C 1.478 against 1.434; below held, 1.635 against 1.733). L7 refuted (the cap arms crash 1, 2 and 4 flights at 1.0 m; t45 without a cap 92-103 as predicted). L8 held (51, 48 and 105 against 55; no arm halves it). L9 held, though t45's +0.001 and +0.002 are inside one se.
  • Reading: "tune at the limit you fly" (7j) is not wrong but buys nothing here that 7k's retrofit does not, and the retrofit also covers the larger reposition retraining misses. The position-kp box edge was binding: with a bounded proportional request the co-design wants about three times the stiffness, and it converts that into a lighter, lower-power airframe at the same tracking. That is the familiar trade of an objective that does not ask for robustness (7i): the lighter vehicle has less torque per unit of gust torque, and its stiffness is safe only while the cap holds. For a builder: a high position gain under a cap is a design that must never fly without the cap, and it should be checked at the centre-of-pressure uncertainty you expect, not only at the nominal one.
  • Limits: the cap value (3.5 m/s²) is 7k's, not swept here; the widened edge is itself binding for two seeds, so the kp the tuner wants may be higher; the objective is lever-1 only with milestone 4's power price, which is what turns stiffness into mass; one simulator, its plant ranges and its start-offset model; six seeds; 64 held-out flights per condition. Simulated; preliminary.
  • Next: the wide box under milestone 7i's lever-aware objective (does it still buy mass, or torque?); sweep the cap (2-6 m/s²) with the wide box; a crash-weighted objective; then map the cap onto a velocity-setpoint limit in a two-loop cascade.

2026-11-15 · Milestone 7k: can a builder remove the 45° crash in the controller, without retuning? (predictions written and committed before experiments/m7k_fixes/run.py was run on the held-out seed; outcome below)

  • Why: milestone 7j's second next step. 7j traced the retrained above designs' lever-0 crashes at a 45° tilt-command limit to the opening transient (a 0.5 m start offset times position kp 10 s⁻² asks a ~27° lean per axis that the 20° training limit had clipped), and its advice was "tune at the limit you fly". A builder with an existing tune would rather fix the controller. Which controller-side change removes the crash, and what does it cost where the designs were trained?
  • Built: four options on poshold.rollout, inert at their defaults (tests/test_poshold.py): tilt_cone (limit the magnitude of the roll/pitch command, keeping its direction, as PX4's MPC_TILTMAX_AIR limits the thrust vector's angle; the simulator so far clipped each axis alone, which lets the combined tilt reach ~55° at a 45° limit: a fidelity gap found while planning this), tilt_rate (slew the roll/pitch reference, rad/s), perr_max (clip the position error entering the proportional term, m; a cascade's velocity-setpoint limit in acceleration form) and sp_rate (the setpoint starts at the vehicle and moves to the hold point at m/s, a speed-limited trajectory; the score still measures distance to the hold point).
  • Plan (experiments/m7k_fixes/run.py, no training): milestone 7i's 48 designs (7e and the 18 retrained above/below designs), 64 held-out episodes (seed 9999), integrator feed-forward, side force on; levers 0, 1, 2; limits 20°, 30°, 45°; fixes base, cone, rate 0.5/1/2 rad/s, perr 0.15/0.25/0.35 m (0.36 m × kp 10 is the 20° lean), ramp 0.5/1/2 m/s. The rate, error bound and ramp speed are swept, not sourced: no autopilot default maps onto this simplified loop. A second block repeats levers 0 and 1 at 45° with the start offset at 1.0 m (a larger reposition). Counts per group; paired change in held-out cost C against base (J's power term does not change).
  • Looked at before writing these: one smoke run (8 episodes, seed 4242, not the held-out seed; 97 s). In it, rate limits of 0.5 and 1 rad/s caused crashes even at 20° (a lagged lean destabilises the stiff position loop), perr and ramp crashed nothing and cost C +0.02 to +0.23, and at a 1.0 m offset base crashed all four sets (7e included) while perr and ramp crashed none and the cone about a third to a half of base. The predictions below are written with that seen.
  • Predictions ("retrained" = 18 designs, 1152 flights; counts at 0.5 m offset unless stated): - K1 (prior 95%): base reproduces 7i/7j's counts: lever 0, 45°: retrained above 24, below 2, 7e above 0, below 1; lever 0, 30°: retrained above 1; lever 2, 20°: retrained above 24, below 72, 7e above 4, below 45. - K2 (prior 55%): the cone cuts the retrained above lever-0, 45° count at least in half. - K3 (prior 85%): perr 0.25 removes all of them (0 of 1152). - K4 (prior 80%): ramp 1 removes all of them. - K5 (prior 75%): rate 0.5 crashes more flights than base at lever 1, 20° (the trained condition) in both geometries. - K6 (prior 65%): perr 0.35 costs less than +0.05 mean C at lever 1, 20° for both groups in both geometries (base C ≈ 1.5). - K7 (prior 60%): none of cone, perr 0.35, ramp 1 cuts the retrained below lever-2, 20° count by more than a quarter (those crashes are 7g's gust-torque tip-overs, not the transient). - K8 (prior 80%): at a 1.0 m offset and 45°, base crashes at least one 7e flight in each geometry, and perr 0.25 and ramp 1 each remove at least 90% of the pooled (all four sets, both levers, both geometries) base count. - K9 (prior 50%): at a 1.0 m offset and 45°, the cone removes at least half of the pooled base count but not all of it.
  • Analysis: counts per group, fix, limit and lever; paired dC per design (mean, se, how many worse). No test beyond counts: six seeds per set, one held-out set.

### Outcome (experiments/m7k_fixes/results.txt, 242 rollout batches of 24 designs × 64 flights, 275 s on 4 cores; figure m7k_fixes.png; simulated)

  • Base reproduced. Every count K1 named comes out exactly (24 / 2 / 0 / 1 at lever 0, 45°; 1 at 30°; 24 / 72 / 4 / 45 at lever 2, 20°).
  • Clipping the position error removes the 45° crash, and so does a setpoint ramp. Retrained above, lever 0, 45°: base 24; perr 0.15/0.25/0.35 m 0 / 0 / 0; ramp 0.5/1/2 m/s 0 / 0 / 0. Over all 48 designs at 45°, lever 1 (the trained lever): base 34 of 3072; perr 0.25 0, perr 0.35 1 (one 7e below flight), every ramp 0. At a 1.0 m start offset the base crashes 935 of 3072 at lever 1 (7e included: 101 and 118 of 384) and perr and ramp still crash none.
  • Their price where the designs were trained (20°, lever 1, mean paired dC over the group, base C about 1.5): perr 0.35 +0.016 to +0.024 (1-1.6%), perr 0.25 +0.060 to +0.076, ramp 2 +0.053 to +0.065, ramp 1 +0.125 to +0.142. Worse in every design (the bound bites in the opening transient, which is part of the score). At 45° the same fixes improve mean C (-0.5 to -0.9 for the retrained above designs) because the crashes go.
  • They also trim the gust-torque crashes at 45°, not at 20°. Lever 2, 45°, all 48 designs: base 244; perr 0.15 163, perr 0.35 206, ramp 0.5 172, rate 2 166. At 20° the retrained below count barely moves (72: cone 68, perr 0.35 74, ramp 1 71): those are 7g's torque-authority tip-overs.
  • A tilt-rate limit is the wrong fix. 0.5 rad/s crashes flights at the trained condition (20°, lever 1: 73 and 69 retrained, 15 and 19 7e, from 0), and 1 rad/s turns 34 crashes at 45° into 515: lagging the lean destabilises a stiff position loop. 2 rad/s cuts 34 to 10 but leaves some and crashes 7e designs that never crashed (5 at lever 0).
  • The cone (PX4's tilt-magnitude limit) changes nothing at the trained offset (24 → 24 at lever 0, 45°; 34 → 34 at lever 1) and removes 59% at 1.0 m (1816 → 703 pooled over both levers). So the per-axis clip the simulator used is not why milestones 7d-7j saw crashes at 45° with a 0.5 m offset: the request that tips them over is under 45° in magnitude. At 30° it matters a little in below (5 → 1 retrained, 5 → 0 7e at lever 1).
  • Verdicts. K1 held. K2 refuted (24 → 24). K3 held (0). K4 held (0). K5 held (73, 15, 69, 19 against 0). K6 held (+0.016 to +0.024). K7 held (68, 74, 71 of 72). K8 held (7e base 97-118 per cell; perr 0.25 and ramp 1 0 of 1816). K9 held (1816 → 703).
  • Reading: the opening-transient crash is the proportional term asking for a lean the tune never saw. Bounding only that term (perr 0.35 m × kp 10 = 3.5 m/s², the 20° lean the designs were trained with) while leaving the derivative and integral terms free to use the full 45° keeps the wind authority and removes every transient crash, for 1-2% of cost where the designs were trained. For a builder with a tune made at a low tilt limit: before raising MPC_TILTMAX_AIR or ANGLE_MAX, cap the position loop's proportional request (in a PX4-style cascade, the velocity setpoint the position error produces) at what the tune saw, or limit the setpoint's speed; do not slow the lean.
  • Limits: the fix values are swept, not taken from an autopilot (this loop is acceleration-form PID, not PX4's P-position / PID-velocity cascade, so MPC_XY_VEL_MAX does not map one to one); one simulator, its plant ranges and start-offset model; six seeds per set; 64 held-out flights per condition. The smoke run on another seed was seen before the predictions. Simulated; preliminary.
  • Next: the same perr bound as a constraint during training at 45° (does the tuner then move kp off its box edge?); map perr onto a velocity-setpoint limit in a two-loop cascade; the crash-weighted objective of 7i.

2026-11-14 · Milestone 7j: why do the retrained above designs diverge with no gust torque at all? (predictions written and committed before experiments/m7j_lever0/run.py was run on the held-out seed; outcome below)

  • Why: milestones 7h and 7i left it unexplained. At lever 0 the centre of pressure sits on the centre of mass: the gust drags but exerts no torque. Yet at the 45° tilt-command limit (PX4's MPC_TILTMAX_AIR default) the retrained above designs diverge at lever 0 on 9 (7h free), 3 (aware TWR=2.0) and 12 (aware TWR free) of 384 held-out flights, where 7e's designs, trained the same way at thrust-to-weight 2, diverge on none, and all pass the linear test with margin 1.25. 7h guessed the larger propellers' slower motors (lag scales as D^1.5). A crash with no disturbance torque at the stock PX4 limit is the kind of finding a builder needs explained before trusting any design here.
  • Plan (experiments/m7j_lever0/run.py, no training): the 48 designs milestone 7i scored (7e, 7h free, aware TWR=2.0, aware TWR free, 6 seeds, both rigid geometries) on the 64 held-out episodes (seed 9999), lever 0, integ feed-forward, side force on, tilt-command limits 30° and 45° (20° base only, as a control), under one probe at a time: base; divergence bound 85°; ideal mixer; position gains x0.5; side force off (couple=False); integrator feed-forward off; every motor's lag x0.9 (about the gap between 7h's propellers and 7e's: nominal lag 15.4 ms against 14.0 ms in the smoke run). The base run records why and when each episode died (before or after the 3 s wind step) and the saturation shares.
  • Looked at before writing these: one smoke run (8 episodes on seed 4242, not the held-out seed): it runs in 13 s and nothing diverged in any set, so it says nothing about the probes; it printed the nominal lags above.
  • Predictions ("retrained" = the 18 above designs of 7h free, aware TWR=2.0 and aware TWR free, pooled; base count at 45° is 24 if J1 holds): - J1 (prior 95%): base reproduces milestone 7i's lever-0 counts exactly (above 45°: 0 / 9 / 3 / 12; 30°: 0 / 0 / 0 / 1; below 45°: 1 / 2 / 0 / 0; 30°: 0 / 2 / 0 / 0; 20°: all 0). - J2 (prior 80%): every retrained above death at 45° is a tilt-over (past 60°), none a 5 m drift. - J3 (prior 70%): at least 2/3 of them happen after the wind step (3 s), when the position loop asks for a large lean against a 6 m/s mean wind. - J4 (prior 55%): position gains x0.5 cut the retrained 45° count at least in half. - J5 (prior 50%): the ideal mixer cuts it at least in half. - J6 (prior 35%): lag x0.9 cuts it at least in half (7h's guess). - J7 (prior 30%): switching the side force off cuts it at least in half. - J8 (prior 40%): switching the integrator feed-forward off cuts it at least in half. - J9 (prior 25%): the 85° bound cuts it at least in half (the 60° bound scoring recoverable excursions). - J10 (prior 60%): no probe makes 7e's above designs diverge at 45° on more than 2 of 384.
  • Analysis: counts per set, probe and limit; why and when; saturation shares in diverged against surviving flights; the plant draws of diverged episodes. No test beyond counts: six seeds, one held-out set.

### Outcome (experiments/m7j_lever0/results.txt, 30 rollout batches, 39 s on 4 cores; post-hoc follow-ups posthoc.txt and offset.txt, under a minute; simulated)

  • Counts reproduced. Base gives milestone 7i's lever-0 counts exactly, in every set and at every limit.
  • Every crash is a tip-over in the opening transient, not in the wind. All 24 retrained above deaths at 45° (and the 1 at 30°, and all 5 below deaths) pass 60° roll or pitch; none drifts. All happen before the wind step: median 0.38-0.39 s into the flight in above (range 0.16-0.79 s), 0.78-1.47 s in below. In diverged flights the mixer scales roll/pitch on a median 74-84% of alive steps and the tilt command sits at its limit on 5-24%; in surviving flights both are 0%. The integrator and its feed-forward are not at their clips (median 0% in both).
  • What removes them (retrained above, 45°, pooled, base 24): ideal mixer 0; position gains x0.5 0; side force off 1; integrator feed-forward off 2; motor lag x0.9 8; 85° divergence bound 23. 7e's above designs diverge on 0 under every probe. Diverged episodes draw slow plants: motor-lag multiplier median 1.91 against 1.59 over all episodes, on 4-7 distinct episodes of 64.
  • Post hoc: it is the recovery from the start offset (offset.txt). The episodes start up to ±0.5 m from the hold point on each axis (poshold.OFFSET). With the offset at 0 or 0.25 m no design in either geometry diverges at lever 0, 45°; at the default 0.5 m the base counts return exactly. Every design's position kp sits at the top of its search box (9.98-10.0 s⁻², posthoc.txt), so a 0.5 m offset asks about 5 m/s², a lean near 27° per axis: the 20° training limit clips that request and 45° passes it. The retrained above designs carry stiffer attitude loops than 7e's (roll/pitch kp 2.33 and 2.41 against 2.13 for 7h free and aware TWR free; aware TWR=2.0, 2.14, has the fewest deaths); across the 24 above designs attitude kp has the largest rank correlation with deaths (+0.57; position kd -0.30).
  • Verdicts. J1 held. J2 held (24 of 24 by tilt). J3 refuted: 0 of 24 after the wind step; the deaths are in the opening transient. J4 held (24 -> 0). J5 held (24 -> 0). J6 held (24 -> 8; 7h's guess explains part, not all). J7 held (24 -> 1). J8 held (24 -> 2). J9 refuted (24 -> 23). J10 held (7e: 0 under every probe).
  • Reading: the lever-0 crash at 45° is a train/deploy mismatch, not a property of the centre of pressure or of the stability margin. Every design was trained at a 20° tilt-command limit, which hid the large-lean regime; at 45° the position loop's opening request passes through, the attitude loop's torque request saturates the hover thrust polytope, and on a slow-motor plant the vehicle tips past 60° within 0.8 s. Any one bit of slack (a gentler outer loop, an ideal mixer, no side force, faster motors) removes most of it, so the retrained designs sit on an edge 7e's do not reach, mainly by their stiffer attitude loops. For a builder: tune at the tilt limit you will fly, or keep the limit at the one you tuned at; ArduPilot's 30° default nearly clears it here (1 of 1152).
  • Limits: the mechanism rests on probes and one offset sweep, all post hoc after J1-J10 except the probes themselves; the correlations are across 24 designs that are not independent draws; the position kp bound (10 s⁻²) is inherited and ASSUMED, and a wider box might push it higher; one simulator, its plant ranges and its start-offset model; 64 held-out flights per design. Simulated; preliminary.
  • Next: train at 45° (or across 20-45°) and check the crash goes; a tilt-rate or acceleration limit on the outer loop as a controller-side fix; widen the position-gain box; then milestone 7i's crash-weighted objective.

### Addendum to 7j: a parallel run's independent check, the onset and a thrust-to-weight re-size (experiments/m7j_lever0/onset_resize.py, onset_resize.txt, 47 s on 4 cores; simulated)

  • Provenance: a second session took the same question from the README's Next steps at the same time, before #15955 was visible, wrote its own script and committed ten predictions of its own (PR #15956, first commit, before its held-out run), and found #15955 merged when it came to land. Its run is kept here only for what #15955 lacks; its predictions that overlap #15955's are not re-scored.
  • Independent reproduction. A separate implementation (designs re-sized with each one's own thrust-to-weight, poshold.rollout called directly) gives #15955's numbers exactly: 9 / 3 / 12 at 45°, lever 0 (above), 1 / 2 / 0 / 0 (below); under the shared probes ideal mixer 0, position gains x0.5 0, side force off 1, feed-forward off 2, lag x0.9 8, 85° bound 23; death times 0.16-0.79 s; mixer scaling on 74-84% of alive steps in diverged flights; 7 distinct episodes.
  • Not thrust-to-weight (new probe). The two thrust-to-weight-free sets (7h free, aware TWR free) re-sized at thrust-to-weight 2.0 with the same frame, propellers, cant and gains: 21 of 21 crashes remain (9 and 12, as in base). The prediction made for it (J7 of #15956, prior 45%: at least half removed) is refuted. So 7h's free thrust-to-weight, and the torque limit it brings, is not what tips these designs over.
  • Onset (new). Summed over the three retrained above sets at lever 0: 1 at 30°, 11 at 35°, 23 at 40°, 24 at 45°; at lever 1: 2, 10, 19, 20. 7e's above designs: 0 at every limit and both levers. The crashes start between 30° and 35° and are nearly all present by 40°; the prediction of a sharp onset above 40° (J10 of #15956, prior 60%: at 40° at most a third of the 45° count) is refuted.
  • Reading: consistent with #15955's mechanism. The re-size rules out the one design variable 7h added, and the onset says that a builder cannot open the limit partway (35°) and stay safe with these gains: about half the crashes are already there.

2026-11-14 · Milestone 7i: a lever-aware objective, train on levers 0-2 with thrust-to-weight fixed and free (predictions written and committed before experiments/m7i_aware/run.py was run at full budget; outcome below)

  • Why: milestone 7h's first next step. 7h found that a tuner given thrust-to-weight as a free coordinate buys a little (2.05-2.55) and spends it on efficiency, and is no more robust at twice the trained lever: nothing in an objective trained at lever 1 asks for robustness to an uncertain centre of pressure. Does putting levers 0-2 in the objective make the co-design buy authority, and does that remove the lever-2 tilt-overs 7g diagnosed?
  • Plan (experiments/m7i_aware/run.py): training score = mean of poshold.score at levers 0, 1, 2, 6 episodes each (18 per candidate against 7e/7h's 16 at lever 1; same training seed per lever). Otherwise milestone 7e/7h's training exactly (integ feed-forward, ki <= ceiling / 1.25 as a constraint, 3200 evaluations, 6 seeds, 2 starts, both rigid geometries). Arms: aware TWR=2.0 (12 coordinates) and aware TWR free (13th coordinate on 1.5-4, start 2.0). Baselines re-scored by the same function: 7e's constrained designs (7e) and 7h's free designs (7h free). Held-out: 64 episodes (seed 9999), levers 0, 1, 2 and 3 (3 is outside every arm's training), tilt-command limits 20°, 30° (ArduPilot ANGLE_MAX default) and 45° (PX4 MPC_TILTMAX_AIR default).
  • Looked at before writing these: one smoke run (64 evaluations, 2 seeds, 2 episodes per lever, scored on seed 4242, not the held-out seed): too short to converge, its numbers are not evidence; it checked the script runs end to end and timed it.
  • Predictions ("diverged" = held-out flights of 384 per set; 7e's lever-2 counts: above 4 at 20°, 12 at 45°; below 45, 55): - A1 (prior 85%): both lever-aware arms diverge on fewer lever-2 flights than 7e at 20° in both geometries. - A2 (prior 50%): aware TWR=2.0 cuts the lever-2 count at least in half at both 20° and 45° in both geometries (7g says gains cannot fix it at thrust-to-weight 2, but the frame geometry, which sets gust torque against torque authority, is still free). - A3 (prior 65%): aware TWR free picks a higher mean thrust-to-weight than 7h free in both geometries (7h: 2.24 above, 2.28 below). - A4 (prior 55%): aware TWR free diverges on 0 lever-2 flights at 20° in both geometries. - A5 (prior 75%): both lever-aware arms pay for it at lever 0, 20°: paired held-out J against 7e below 0 in both geometries. - A6 (prior 50%): that price is smaller than the retrofit's in 7h: aware TWR free minus 7e at lever 0, 20° above -0.128 in both geometries (7h's retrofit max: -0.144 above, -0.128 below). - A7 (prior 40%): at lever 3 (outside all training), aware TWR free diverges on at most half 7e's count at 20° in both geometries. - A8 (prior 70%): for every set, the lever-2 count at 30° lies between its counts at 20° and 45° (inclusive).
  • Analysis: paired per-seed differences against 7e (mean, se, sign count), divergence counts per set and condition; thrust-to-weight, torque limit per newton of weight, geometry. No test beyond counts with six seeds.

### Outcome (experiments/m7i_aware/results.txt, 48 CMA-ES runs of 3200 evaluations and 48 designs re-scored, about 2 h 20 min on 4 cores; the first launch hit the session's 2-hour background limit after 44 of 48 runs, which were cached, and the resumed run reports only its own 751 s; post-hoc posthoc.txt; simulated)

  • Baselines reproduced. 7e's lever-2 counts (above 4 at 20°, 12 at 45°; below 45, 55) and 7h's free counts come out exactly as in 7e/7h.
  • below: the lever-aware objective roughly halves the crashes at no cost at the design lever. Diverged at lever 2, 20°: 45 (7e) -> 20 (aware TWR=2.0) and 14 (aware TWR free); at lever 3, outside all training, 159 -> 87 and 86. Held-out J at lever 2, 20°: +2.18 and +2.47 against 7e (better in 6 of 6 seeds each); at lever 0, 20°: +0.011 and +0.012 (5 and 4 of 6), at lever 1: -0.026 and -0.022 (1 and 0 of 6).
  • above: no gain. Lever 2, 20°: 4 (7e) -> 7 and 8; lever 3: 68 -> 66 and 75. 7e's above designs crashed on 4 of 384 lever-2 flights, so 6 training episodes per lever see almost no crashes and give the tuner little to push against.
  • Thrust-to-weight free adds nothing over thrust-to-weight 2. The lever-aware tuner picks 2.16-2.47 (above, mean 2.32) and 2.07-2.67 (below, mean 2.30), within 0.08 of 7h's lever-1 choices (2.24, 2.28); its counts match the fixed-2 arm within noise (below 14 against 20 at 20°, 23 against 22 at 30°, 29 against 28 at 45°).
  • What the tuner buys instead is mass (post hoc). At fixed thrust-to-weight a heavier design gets proportionally bigger motors and torque limit, while the gust torque is set by the frame's area and lever, not its mass. Across all 24 designs per geometry (four arms, six seeds), the rank correlation of mass with lever-3 divergences at 20° is -0.87 (above) and -0.94 (below); of torque limit -0.81 and -0.87; of thrust-to-weight +0.08 and -0.02. The lever-aware below designs weigh 1.49-1.52 kg against 7e's 1.23 kg (hover power 428 W against 316 W). Seed 3 of all four lever-aware runs landed on a ~2.0 kg design (hover power 656-678 W) that diverges on 0 lever-2 and 4-5 lever-3 flights; the others stay at 1.3-1.7 kg. The objective still prices hover power, so most seeds stop short.
  • 30° (ArduPilot's default). Every set's lever-2 count at 30° sits between its 20° and 45° counts; 7e's above designs diverge on 9 of 384 there, below on 47. At 45°, aware TWR free above diverges at lever 0 on 12 of 384 (7h free 9, aware TWR=2.0 3, 7e 0): all three retrained above arms show the lever-0 divergence at the loose limit that 7e's designs do not; still not explained.
  • Verdicts. A1 refuted: held in below, not in above (7 and 8 against 4). A2 refuted: below 45 -> 20 halves at 20°, but 55 -> 28 at 45° misses half (27.5) by one flight; above worse. A3 held as worded, by margins (0.08, 0.02) too small to mean anything. A4 refuted (8, 14). A5 refuted: only aware TWR free above pays at lever 0, 20° (-0.007, se 0.008); the other three paired means are above 0. A6 held (-0.007, +0.012, both above -0.128). A7 refuted (above 75 against 68; below 86 against 159, more than half). A8 held in all 8 sets.
  • Reading: a lever-aware objective helps where the lever-1 designs were fragile (below) and does it by choosing a heavier vehicle, not more thrust-to-weight: on this sizing model the torque limit scales with mass at fixed thrust-to-weight and the gust torque does not. 7h's retrofit removed the crashes for the same reason (2.86 kg). For a builder facing an uncertain centre of pressure, the useful quantity is roll/pitch torque against the frame's gust torque (area times lever), and on an endurance-sized airframe thrust-to-weight is a poor handle on it. The design question for milestone 3 becomes torque per unit gust torque, not thrust-to-weight.
  • Limits: the mass result is post hoc and correlational across designs that are not independent draws; inherited motor-mass and sizing fits decide how mass, torque limit and hover power move together, and the gust model makes gust torque mass-independent by construction; six seeds and 6 training episodes per lever (a lever-2 crash rate near 1% gives the above tuner little signal); one simulator; 64 held-out flights per condition. Simulated; preliminary.
  • Next: an objective that weights the crash rate directly (or more lever-2 training episodes) so above gets a signal; a torque-per-gust-torque design variable checked against the sizing; the retrained above designs' 45° lever-0 divergence with 7g's diagnostics; ten seeds for the paper figures.

2026-11-14 · Milestone 7h: thrust-to-weight as a design variable, scored at twice the trained lever (predictions written and committed before experiments/m7h_twr/run.py was run at full budget; outcome below)

  • Why: milestone 7g's first next step. 7g found the residual lever-2 divergence is a tilt-over that starts in the mixer: the gust torque outgrows the differential thrust the hover polytope allows, and about 25% more roll/pitch authority removes it. Its advice to a builder was "buy thrust headroom (thrust-to-weight above 2)". Every design so far was sized at thrust-to-weight 2 (aero.TWR). Does the co-design choose headroom when it may, and can a builder add it after the fact?
  • Built: an optional twr on poshold.score and linstab.design_params (default None keeps 2.0 bit-for-bit; tests/test_poshold.py). Motor mass, cost and torque limits follow it through aero.size, and so does hover power through the mass.
  • Plan (experiments/m7h_twr/run.py): arm free retrains milestone 7e exactly (integ feed-forward, ki <= ceiling / 1.25 as a constraint, 3200 evaluations, 16 episodes at lever 1, 6 seeds, 2 starts, both rigid geometries) with a 13th coordinate for thrust-to-weight on 1.5-4 (start 2.0). Arm retrofit keeps 7e's constrained designs and gains and re-sizes them at 2.0 and at the largest thrust-to-weight (0.05 grid) whose sizing still closes. Held-out: 64 episodes (seed 9999), levers 0, 1, 2, tilt-command limits 20° and 45°.
  • Looked at before writing these: two 64-evaluation smoke runs (2 seeds, 4 episodes, scored on seed 4242, not the held-out seed; too short to converge). The first planned retrofit at 2.5 and 3.0 and found the 15-minute endurance closure runs away to the 20 kg clamp. A direct check on the 12 designs (no simulation): the sizing closes up to thrust-to-weight 2.10-2.15 for above and 2.25-2.50 for below; at 2.25 an above design's hover power goes from 450-560 W to about 20 kW. The co-designed airframes sit on the edge of the mass spiral (heavier motors -> more battery -> more thrust needed), so the retrofit arm was changed to "largest that closes" before any held-out run.
  • Predictions: - H1 (prior 90%): retrofit at 2.0 reproduces milestones 7e/7g's lever-2 counts exactly (above: 4 at 20°, 12 at 45°; below: 45 at 20°, 55 at 45°). - H2 (prior 55%): retrofit max cuts the lever-2 divergences at least in half in both geometries at both limits (more differential thrust per unit gust torque). - H3 (prior 75%): retrofit max costs held-out J at lever 0 (20°) in both geometries (paired mean below 0): the heavier vehicle's hover power price outweighs any tracking gain. - H4 (prior 45%): at least one retrofit max design is linearly unstable at some plant corner (worst radius >= 1): the PID outputs are shares of the torque limit, so bigger motors raise the loop gain of a design trained at its ki bound. - H5 (prior 70%): free picks thrust-to-weight below 2.0 in at least 4 of 6 seeds in each geometry: the objective prices hover power and trains only at lever 1, where 7e's designs did not diverge. - H6 (prior 60%): free diverges at lever 2, 20°, on more held-out flights than retrofit 2.0 in at least one geometry. - H7 (prior 55%): free beats retrofit 2.0 on held-out J at lever 0, 20° (paired mean above 0) in both geometries.
  • Analysis: paired per-seed differences against retrofit 2.0 (mean, se, sign count), divergence counts per set. No test beyond counts with six seeds.

### Outcome (experiments/m7h_twr/results.txt, 24 CMA-ES runs of 3200 evaluations and 36 designs re-scored, 2241 s on 4 cores; simulated)

  • The co-designed airframes have almost no room for bigger motors. On 7e's airframes the 15-minute endurance sizing closes only up to thrust-to-weight 2.10-2.15 (above) and 2.25-2.50 (below). There, mass goes from 1.61 to 2.86 kg (above), hover power from 486 to 1157 W and cost from $134 to $248. The torque limit nearly doubles (0.98 to 1.74 N m) because the heavier vehicle hovers on more thrust, not because of the extra 0.1-0.15 of thrust-to-weight alone.
  • Retrofit at the edge removes the lever-2 crashes, at a price. Diverged held-out flights at lever 2, 20° / 45°: above 4 / 12 -> 0 / 0, below 45 / 55 -> 0 / 2. Held-out J at lever 0, 20°, paired against thrust-to-weight 2: -0.144 (se 0.016, worse in 6 of 6) above, -0.128 (se 0.017, 6 of 6) below, almost all of it the hover-power price. At lever 2 the same designs gain +0.16 to +4.58.
  • Retraining with headroom free picks more of it and spends it elsewhere. All 12 seeds choose thrust-to-weight above 2 (above 2.09-2.37, below 2.05-2.55) with slightly larger propellers (D 0.129 m against 0.121 m above) and lower hover power (460 W against 486 W above; 312 W against 316 W below). Their torque limit stays where it was (0.995 against 0.981 N m above; 0.804 against 0.821 below). Held-out J at lever 0, 20°: +0.006 (se 0.002, better in 5 of 6) above, +0.001 (se 0.004, 2 of 6) below.
  • And they are no more robust out of training. Lever 2, 20°: above 9 diverged (against 4), below 38 (against 45). At 45° the retrained above designs diverge even at lever 0 (9 of 384) and lever 1 (10), where 7e's designs diverged on none; every one passes the linear test with margin 1.25 (worst radius at most 0.999998, ceiling multiples 1.25-1.36). Not explained here; the larger propellers' slower motors (lag scales as D^1.5, about +10%) are a candidate, not a finding.
  • Verdicts. H1 held (4, 12, 45, 55 reproduced). H2 held (above 4 -> 0 and 12 -> 0; below 45 -> 0 and 55 -> 2). H3 held (-0.144, -0.128). H4 refuted: no retrofit design is linearly unstable (0 of 12), and the ceiling multiples rise slightly (1.27-1.30 -> 1.28-1.33 above): the heavier vehicle's inertia grows with its torque limit. H5 refuted: 0 of 12 seeds pick below 2.0. H6 held in above (9 against 4), not in below (38 against 45). H7 held as worded (both paired means above 0), but below's +0.001 is within one se.
  • Reading: "buy thrust headroom", 7g's advice, is right about what removes the crash and wrong as a recipe. On an airframe sized to its endurance target the headroom comes with a mass spiral: 0.1-0.15 more thrust-to-weight costs 78% more mass and 2.4 times the hover power. A tuner given headroom as a free variable buys some, then uses it on efficiency, because nothing in its objective asks for robustness to a larger gust lever. Robustness to an uncertain centre of pressure has to be in the objective (train on levers 0-2), not left to the design space.
  • Limits: inherited motor-mass and sizing fits (milestone 3's catalogue data would replace them; the closure's cliff depends on them); the retrofit keeps the 15-minute endurance target, where a builder might accept shorter flights instead; one simulator, its gust model and plant ranges; six seeds; 64 held-out flights per design and condition. The propeller-lag reading of the 45° divergence is a guess. Simulated; preliminary.
  • Next: train on levers 0-2 (lever-aware objective) with thrust-to-weight free and see whether the tuner then buys authority; probe the retrained above designs' 45° divergence at lever 0 with 7g's diagnostics; a retrofit that keeps the battery and gives up endurance instead of mass.

2026-11-14 · Milestone 7 started: the paper draft, with the stability section written and every quoted number checked

  • Why: milestone 7f's third next step ("the paper's stability section with this curve as its figure"); the programme has eight milestones of results and no paper. A sibling run holds milestone 7g (lever-2 divergence, PR #15752); this run took the draft so the two do not collide.
  • Built: paper/DRAFT.md (abstract, sections I–V and VII–VIII outlined, section VI on the feed-forward's stability written in full with Table VI-1 from milestone 7f), paper/references.bib (21 entries; 18 DOIs and arXiv ids ok* from cite_check.json, Bartlett et al. 1988 not machine-checked: the proxy refused Crossref and doi.org), paper/check_numbers.py (35 checks of quoted text against regular expressions on the results files) and tests/test_paper_numbers.py.
  • Found while writing: the tilt-command limit the simulator uses (20°) is the minimum PX4 allows for MPC_TILTMAX_AIR, whose default is 45°; ArduPilot's ANGLE_MAX defaults to 30°. Read from multicopter_position_control_limits_params.yaml (PX4-Autopilot, main) and AC_AttitudeControl.cpp (ArduPilot, master) on 2026-11-14. So milestones 7d–7f's 45° rows are the stock PX4 condition, and their headline (13 of 384 diverged without the margin, 0 with it) is the deployment-relevant one. Not a new experiment; a reading of existing results against a sourced setting.
  • Limits: no new numbers; the draft inherits every limit in README.md. Six seeds per arm where the proposal asks ten for paper figures.
  • Next: score milestones 6–7f's designs at 30° (ArduPilot's default) too; write sections I–V out; ten seeds and bootstrap intervals for the paper figures; LaTeX in the RA-L template.

2026-11-14 · Milestone 7g: what makes the margin-constrained designs diverge at twice the trained lever? (predictions written and committed before experiments/m7g_lever2/run.py was run on the held-out seed; outcome below)

  • Why: milestone 7f's first next step. From 0.8x the linear ceiling down to 0.6x, the above designs still diverge on 4 of 384 held-out flights at lever 2 with the 20° tilt-command limit and 12-13 at 45°; milestone 7e's below designs, where the margin does not bind, diverge on 45. A tighter margin does not move these counts, so something other than the integrator's linear stability drives them. Candidates: the attitude-priority mixer saturating, a gust torque beyond the motors' authority, the 60° tilt bound scoring a recoverable excursion as a crash, or the outer position loop.
  • Built: four more poshold.rollout probes that leave the default rollout bit-for-bit unchanged (tests/test_poshold.py): dead_tilt_deg (the divergence bound), mixer_ideal (the clipped request delivered with no polytope scaling), torque_room (every torque limit scaled), pos_scale (position-loop gains scaled); diag now also records why each episode died (tilt or position) and the gust torque against the plant's torque limit.
  • Plan (experiments/m7g_lever2/run.py, no training): milestone 7b's unconstrained integ retrained, 7e's constrained (M = 1.25, both geometries) and 7f's M = 1.43 and 1.67 (above) designs, 6 seeds each, on the 64 held-out episodes (seed 9999), lever 2, tilt-command limit 20° and 45°, under one probe at a time: base, dead bound 85°, ideal mixer, ideal mixer with 2x torque room, position gains x0.5.
  • Looked at before writing these: an 8-episode smoke run on seed 4242 (not the held-out seed). One death in 384 design-episodes (below, unconstrained, 45°), by tilt, with the mixer scaling roll/pitch on 81% of its alive steps; the gust torque peaked at about 0.4 of the plant's torque limit in every set. So a plain lack of authority against the gust looks unlikely; the predictions below take that into account.
  • Predictions (pooled over the above M >= 1.25 sets unless stated; "cut" = fewer diverged episodes than base): - G1 (prior 95%): base reproduces milestone 7f's and 7e's counts exactly (above 20°: 15 unconstrained, 4/4/4; 45°: 45, 12/13/12; below 20°: 42 unconstrained, 45 constrained). - G2 (prior 75%): authority is not the cause: the gust torque exceeds the plant's torque limit at some alive step in at most 20% of diverged episodes. - G3 (prior 70%): at least 2/3 of diverged episodes die by tilt (past 60°), not by drifting 5 m. - G4 (prior 60%): the ideal mixer cuts the pooled count by at least half at both tilt limits. - G5 (prior 65%): the ideal mixer with 2x torque room cuts it by at least 75% at both limits. - G6 (prior 30%): the 85° dead bound cuts it by at least half (the 60° bound mostly scores recoverable excursions). - G7 (prior 35%): halving the position gains cuts it by at least half. - G8 (prior 50%): in below (constrained, 20°) the same probe that cuts above most also cuts it by at least half; the cause is shared.
  • Analysis: counts of diverged episodes per set and probe; medians of the saturation shares in diverged against surviving episodes; the plant draws of the diverged episodes. No test beyond counts: 6 seeds, one held-out set.

### Outcome (experiments/m7g_lever2/results.txt, 20 rollout batches, 31 s on 4 cores; post-hoc follow-up headroom.txt, under 2 minutes; simulated)

  • Every lever-2 divergence is a tilt-over. All 49 constrained above deaths (12 at 20°, 37 at 45°), all 60 unconstrained above and all 100 constrained below deaths pass 60° roll or pitch; none drifts 5 m first.
  • The mixer is where they start. In diverged flights the attitude-priority mixer scales roll/pitch on a median 6-51% of alive steps; in surviving flights on 0%. With an ideal mixer (the clipped request delivered, no polytope), constrained above deaths fall from 12 to 4 (20°) and 37 to 8 (45°); below from 45 to 15 and 55 to 18. With an ideal mixer and 1.25x torque room, above falls to 0 at both limits (unconstrained: 0 and 1); below needs 1.5x to reach 0 (6 at 1.25x).
  • Authority, not stability. The gust torque at lever 2 peaks at a median 0.87-1.13 of the plant's (randomised) torque limit in diverged constrained above flights against 0.50 in survivors, and exceeds the limit at some alive step in 24 of 49 (11 of 12 at 20°). Surviving flights see it above the limit in 1-4%.
  • The other suspects are weak. Dead bound 85°: 12 -> 8 and 37 -> 29 (the 60° bound is not mainly scoring recoverable excursions). Position gains x0.5: 12 -> 10 at 20° but 37 -> 12 at 45° (a gentler outer loop asks less tilt and so less torque when the tilt command is loose).
  • Verdicts. G1 held (every count reproduced). G2 refuted: the gust exceeds the limit in 24 of 49 diverged flights, not at most 20%; the smoke run's 0.4 peak was a calm seed. G3 held (all by tilt). G4 held (12 -> 4, 37 -> 8). G5 held (0 and 0). G6 refuted (-33%, -22%). G7 split: pooled over both limits 49 -> 22 (held as worded), but at 20° only 12 -> 10. G8 held: the ideal mixer with 2x room removes all 100 below deaths and the ideal mixer alone two thirds.
  • Follow-up (experiments/m7g_lever2/headroom.py, written after the results above were read). Airmode (the mixer may shift the collective, allocation.mix_priority(mode="free")) does not help: constrained above 12 -> 19 and 37 -> 40, below 45 -> 39 and 55 -> 52. At hover the collective already sits mid-range (thrust-to-weight 2), so shifting it buys nothing. More torque behind the real mixer barely helps either (airmode with 1.5x torque: above 11 and 30, below 28 and 38): the hover thrust polytope, set by the motors' maximum thrust, is the binding limit. An ideal mixer with 1.25x torque, a stand-in for about 25% more roll/pitch authority than the polytope gives, removes every constrained above divergence.
  • Reading: the residual lever-2 divergence is not a stability-margin problem and no gain or margin will fix it; the vehicle runs out of differential thrust when the gust torque doubles. For a builder: if the frame's centre of pressure may sit twice as far above the centre of mass as designed, buy thrust headroom (thrust-to-weight above 2), not a different tuning or airmode. That makes milestone 3 (motor headroom free under the polytope mixer) the next design question.
  • Limits: probes, not redesigns: torque room scales limits without resizing the motors, mass or power, and the ideal mixer is not realisable; one simulator, its assumed gust model and plant ranges, six seeds, one held-out set of 64 flights per design; lever 2 is outside training. The headroom and airmode checks are post hoc. Simulated; preliminary.
  • Next: retrain with thrust-to-weight free (milestone 3's question) and score at lever 2; a lever-aware design (train on levers 0-2); the paper's stability-and-authority section with 7c-7g as its figures.

2026-11-13 · Milestone 7f: sweep the linear stability margin used as a training constraint (predictions written and committed before experiments/m7f_sweep/run.py was run at full budget; outcome below)

  • Why: milestone 7e's first next step. 7e tested one margin (ki <= 0.8x the linear ceiling, M = 1.25) and found it costs about 0.7% of held-out cost and removes every divergence a 45° tilt limit exposes. A builder needs the curve, not one point: how much margin is enough, and what does more cost?
  • Setup (experiments/m7f_sweep/run.py): milestone 7e's training exactly (12-coordinate integ arm, 3200 CMA-ES evaluations, 16 training episodes, 6 seeds, 2 starts) for the centre-of-pressure-above geometry only (7e: the constraint does not bind below), at M = 1.0 (only "stable at every plant corner at its own ki"), 1.11 (0.9x), 1.43 (0.7x), 1.67 (0.6x). M = 1.25 is 7e's designs and the unconstrained arm is 7b's integ retrained; all seven arms are re-scored by one function on the 64 held-out episodes (seed 9999), lever 0, 1, 2, at tilt-command limits 20° and 45°.
  • Looked at before writing these: a 64-evaluation smoke run (2 seeds, 4 training episodes, scored on seed 4242, not the held-out seed; too short to converge, its numbers are not evidence). It showed that 86-87% of the first candidates are rejected at M >= 1.43, so the start point is far from feasible there.
  • Predictions: - F1 (prior 85%): at every M, every seed's best design is feasible (best training score above -100). - F2 (prior 75%): the mean share of training candidates rejected rises strictly with M across 1.0, 1.11, 1.25 (7e: 0.13), 1.43, 1.67. - F3 (prior 70%): the constraint binds for every M >= 1.11: ceiling multiple in [M, M + 0.15] in at least 5 of 6 seeds. - F4a (prior 55%): the price is monotone: mean held-out J at 20°, lever 0 does not increase from one arm to the next across unconstrained, 1.0, 1.11, 1.25, 1.43, 1.67. - F4b (prior 60%): at M = 1.67 the paired price at 20°, lever 0 is still at least -0.05 (7e at 1.25: -0.011). - F5 (prior 60%): at 45°, lever 0, the M = 1.0 designs diverge on at least 1 of 384 flights (they sit on the boundary, like the unconstrained designs' 13), and every arm with M >= 1.11 diverges on 0. - F6 (prior 45%): at 20°, lever 2 (outside training), total divergences do not increase with M across 1.0, 1.11, 1.25 (7e: 4), 1.43, 1.67.
  • Analysis: per-arm means, paired per-seed differences against unconstrained (mean, se, sign count), divergence counts. No test beyond the counts with six seeds.

### Outcome (experiments/m7f_sweep/results.txt, 48 new CMA-ES runs of 3200 evaluations, 4055 s on 4 cores; 42 designs re-scored on 64 held-out flights at 6 conditions; simulated)

  • The optimiser sits on whatever bound it is given. Ceiling multiples: M = 1.0: 1.00-1.07; 1.11: 1.12-1.14 (5 seeds; seed 2 dropped the integrator, ki/kp 0.03); 1.43: 1.43-1.53; 1.67: 1.67-1.74. ki/kp falls from 3.6-4.3 (M = 1.0) to 2.4-3.2 per s (M = 1.67); cant 31-34° and hover power 475-486 W in every arm.
  • The price is flat past 0.9x. Held-out J at 20°, lever 0, paired against unconstrained: M = 1.0 -0.001 (se 0.004), 1.11 -0.012 (0.010), 1.25 -0.011 (0.003), 1.43 -0.010 (0.003), 1.67 -0.012 (0.006). Position RMS 0.105 m in every arm. Cutting ki from 0.8x to 0.6x of its ceiling costs nothing measurable; the other gains take up the slack.
  • Robustness has a knee at 0.8x. Diverged held-out flights (of 384), unconstrained / 1.0 / 1.11 / 1.25 / 1.43 / 1.67: 45° limit, lever 0: 13 / 9 / 3 / 0 / 0 / 0; 45°, lever 1: 13 / 9 / 2 / 0 / 0 / 0; 20°, lever 2: 15 / 12 / 9 / 4 / 4 / 4; 45°, lever 2: 45 / 32 / 22 / 12 / 13 / 12. Below 0.8x no count improves further.
  • Training cost. Share of candidates rejected 0.11 / 0.10 / 0.13 (7e) / 0.14 / 0.15; wall time per run is no worse than unconstrained.
  • Verdicts. F1 held (lowest best -1.69). F2 refuted, narrowly (0.11 at M = 1.0 against 0.10 at 1.11). F3 held (5 of 6 at 1.11, 6 of 6 at 1.25, 1.43 and 1.67). F4a refuted: the price is not monotone, it steps from 0.000 to about -0.011 between M = 1.0 and 1.11 and stays there. F4b held (-0.012 against a -0.05 bar). F5 split: M = 1.0 diverges (9 of 384), but so does 1.11 (3 of 384). F6 held (12, 9, 4, 4, 4).
  • Reading: for a builder adding the linear check to a tuning loop, 0.8x the ceiling is the knee: 0.9x still leaves 3 of 384 flights diverging once the tilt limit is lifted, and anything from 0.8x down to 0.6x gives the same robustness at the same cost (about 0.7% of held-out J). "Stable at every plant corner" alone (M = 1.0) is not enough: it removes a third of the loose-limit divergences, not all of them. The residual lever-2 divergence at a 45° limit (12 of 384 at every M >= 1.25) is not a margin effect.
  • Limits: one geometry (above), six seeds, one simulator and its assumed plant ranges; the margin is on ki only (the rest of the loop only has to be stable, no margin); 64 held-out flights per design per condition. Simulated; preliminary.
  • Next: what drives the residual lever-2 divergence (saturation in the mixer, the 60° tilt bound, or the outer position loop); a margin on the whole loop (spectral radius bound below 1) instead of on ki alone; the paper's stability section with this curve as its figure.

2026-11-13 · Milestone 7e: train with milestone 7c's linear stability margin as a CMA-ES constraint (predictions written and committed before experiments/m7e_margin/run.py was run at full budget)

  • Why: milestone 7c's second next step. Left alone, CMA-ES put the integ designs' attitude integrator gain on the linear stability boundary (ceiling multiples 0.88-1.11), with nothing between the design and the edge of the plant set. 7c's advice to a builder, "stay at about 0.8x the ceiling", was checked only by rescaling designs after training. Here the margin is a training constraint: does a design trained inside it lose cost, and does it diverge less?
  • Setup (experiments/m7e_margin/run.py): milestone 7b's integ arm (12 coordinates; milestone 6's CMA-ES, 3200 evaluations, 16 training episodes, 6 seeds, 2 starts; both rigid geometries). A candidate is feasible if linstab.worst_radius (both axes, the eight wide plant corners) is below 1 at its own ki and at 1.25x its ki, i.e. ki <= 0.8x ceiling and the rest of the loop stable at every corner. Infeasible candidates are not simulated and score -100 - 10 min(rho - 1, 10). Held-out: 64 episodes, seed 9999, lever 0, 1, 2, side force on; compared, seed for seed, with milestone 7b's unconstrained integ retrained designs re-scored by the same function.
  • Looked at before writing these: a 64-evaluation smoke run (training seeds only, held-out seed untouched; too short to converge, its numbers are not evidence); 4% of uniform random designs are feasible and the CMA-ES start point is not (rho 1.029); the slowest mode of every design sits at rho 0.9997-1.0002, so feasibility is decided by a slow mode; of the 7b designs cut to 0.8x ki, 10 of 12 would pass the test (above seeds 1 and 5 fail; below seed 1 is unstable at a corner even with ki cut).
  • Predictions: - E5 (prior 90%): every seed's best design is feasible (no best score at or below -100). - E1 (prior 70%): the constraint binds in above: ceiling multiple between 1.25 and 1.45 in at least 4 of 6 seeds. - E2 (prior 60%): the price is small: mean paired held-out J (constrained minus unconstrained) at lever 0 is at least -0.05 in both geometries. - E3 (prior 70%): no held-out divergence at lever 1 in any of the 12 constrained designs (unconstrained: 2, below seed 1). - E4 (prior 35%): at lever 2 (outside the training range) the constrained designs diverge on fewer held-out episodes in total than the unconstrained ones, in both geometries (unconstrained: 15 above, 42 below). Low prior: milestone 7c found divergence there is saturation-limited, which the linear test does not see.
  • Analysis: paired per-seed differences, mean and se, counts of diverged episodes; no test beyond the counts with six seeds.

### Outcome (experiments/m7e_margin/results.txt, 24 CMA-ES runs of 3200 evaluations, about 34 minutes on 4 cores; follow-up tilt.txt, under 2 minutes; simulated)

  • Training stayed feasible. 13% of above and 8% of below training candidates were rejected by the linear test (range 4-21%); every seed's best design is feasible (lowest best score -3.36). The test is cheap: the 24 runs took about 34 minutes against milestone 7b's 50.
  • The constraint binds in above and the optimiser sits on it. Ceiling multiples 1.27, 1.28, 1.30, 1.28, 1.26, 1.30 (unconstrained: 0.88-1.11); ki/kp 3.3-3.9 per s against 3.7-4.4. Cant 30-33° against 32-36°. In below, ki/kp stays 0.09-0.24 and ceiling multiples 2.1-4 (the ki margin does not bind there; the constraint acts only by requiring the whole loop to be stable at every corner, which below seed 1's unconstrained design was not).
  • Price. Held-out J, constrained minus unconstrained, paired by seed: above -0.011 at lever 0 (se 0.003, worse in 6 of 6, about 0.7% of J), -0.009 at lever 1; below -0.000 at lever 0, +0.198 at lever 1 (se 0.199; seed 1's two divergences gone). Position RMS unchanged to the third decimal in above (0.105 m).
  • Divergence. Lever 0 and 1: none in any of the 12 constrained designs (unconstrained: 2, below seed 1, lever 1). Lever 2 (outside training): above 4 of 384 against 15 (fewer in 6 of 6 seeds; J +0.60, se 0.22, better in 6 of 6); below 45 against 42 (J -0.15, se 0.40).
  • Verdicts. E5 held. E1 held (6 of 6 in [1.25, 1.45]). E2 held (-0.011 and -0.000 against a -0.05 bar). E3 held (0 of 768 at lever 1). E4 refuted: fewer lever-2 divergences in above (4 against 15) but not in below (45 against 42).
  • Follow-up (experiments/m7e_margin/tilt.py, written after the results above were read; milestone 7d's next step). Tilt-command limit 20/30/45/59°, lever 0: above unconstrained diverge on 0/2/13/13 of 384 flights (position RMS 0.105 -> 0.257 m at 45°), constrained on 0/0/0/0 (0.103-0.105 m); below unconstrained 0/2/2/2, constrained 0/0/1/1. Lever 1, 45°: above 13 against 0; below 6 against 6. The 13 at 45° match milestone 7d's 3-4% for designs on the ceiling.
  • Reading: for the above geometry, where the integrator carries the gust torque, the 0.8× linear margin as a training constraint costs about 0.7% of held-out cost and removes every divergence that a looser tilt limit or a doubled centre-of-pressure lever exposes in designs tuned onto the boundary. It is a cheap certificate a builder can add to any tuning loop: one 11-state eigenvalue check per candidate at eight plant corners. Where the integrator does little (below), the constraint changes little, and the lever-2 divergence there has another cause.
  • Limits: same simulator and plant ranges as milestones 6-7d; six seeds; one margin (0.8×); the lever-2 and tilt-limit scorings are single held-out sets (64 flights per design); the tilt-limit check is post hoc. Simulated; preliminary.
  • Next: a margin sweep (0.6-1.0×) to trace cost against robustness; the same constraint for the commanded-torque feed-forward; what drives below's lever-2 divergence; the describing-function analysis 7d proposed; the paper's stability section.

2026-11-13 · Milestone 7d: what keeps a linearly unstable integrator feed-forward bounded, the integrator clip or the episode length? (predictions written and committed before experiments/m7d_clip/run.py was run)

  • Why: milestone 7c found its linear test necessary but not sufficient in integrator mode. Milestone 6b's above designs at 1.25× their ki ceiling are linearly unstable at 80–100% of held-out plants (growth 0.50–0.66 per s) yet none diverged, and the bolted-on designs (growth 3.9–7.4 per s) diverged on only 11–44% of episodes. 7c's untested explanation: the attitude integrator's clip (±0.3 rad s) caps the feed-forward. A rival explanation it did not rule out: a growth of 0.5 per s needs about 9 s to amplify 100×, longer than the 6 s episode, so the unstable mode may simply not have time.
  • Built: poshold.rollout takes three probes that leave the default rollout bit-for-bit unchanged (tests/test_poshold.py): integ_clip (the attitude integrator clip), ffi_clip (the ±1 clip on ki·integ in the integrator feed-forward), steps (episode length; same seed keeps the first draws and the four yaw signs), and diag, which records per episode the share of alive steps each saturation is active (integrator at clip, feed-forward at clip, tilt command at 20°, mixer scaling roll/pitch) and the time of divergence.
  • What was looked at before writing these: a smoke run on 4 episodes of seed 4242 that printed array shapes only. No 7d number on the held-out set.
  • Plan: experiments/m7d_clip/run.py. Milestone 7c's 24 above integrator-mode design-conditions with the feed-forward on (bolted-on 6b+integ, 6b at 1.25× and 0.8× its ceiling, integ retrained; 6 seeds each), the 64 held-out episodes (seed 9999), clip arm {base: 0.3 and 1; noint: integrator clip removed; none: both removed} × horizon {6 s, 18 s} × lever {0, 1}. Also the 11-state linear map iterated from each episode's initial offset alone (no gust, wind or noise) to the first time |p| > 5 m or 60° tilt.
  • Predictions (lever 0 unless stated; "unstable" = the linear model at the episode's own plant): - V1 (reproduction): base, 6 s reproduces 7c's per-design divergence counts exactly at both levers. - V2 (7c's conjecture, fast instability): removing the integrator clip at least doubles the bolted-on designs' divergences at 6 s (7c: 84 of 384). - V3 (7c's conjecture, slow instability): with the integrator clip removed and 18 s episodes, the 1.25× designs diverge on at least half the episodes the linear model calls unstable (pooled over 6 designs). - V4 (the rival, horizon alone): with the clip kept, 18 s episodes make the 1.25× designs diverge on at most 10% of all 384 episodes (if the clip is what bounds them, more time does not help). - V5 (control): the 0.8× and retrained designs diverge on at most 1 of 64 episodes per design in every clip arm and horizon. - V6 (mechanism): at base, 6 s, the roll/pitch integrator sits at its clip on at least 5% of alive steps (median episode) for the 1.25× designs, and on more for the bolted-on designs than for the 0.8× designs. - V7 (linear, offset only): for the 1.25× designs the linear model reaches |p| > 5 m or 60° within 18 s in at least half of its unstable episodes and within 6 s in at most 10%. - V8: removing the feed-forward clip as well (none) diverges at least as often as noint, pooled, in every set and horizon.
  • Analysis: counts per design and pooled, median divergence time, median saturation shares. No test beyond the counts.

### Outcome (experiments/m7d_clip/results.txt, 24 design-conditions × 3 clip arms × 2 horizons × 2 levers × 64 held-out episodes, about 3 minutes on 4 cores; simulated)

  • Correction to the plan's wording: the 18 s episodes keep the seed, the 64 plant draws and the four yaw signs, but the gust, noise, initial offsets and wind-step signs are drawn afresh (the random arrays change length). They are new flights on the same plants, not the 6 s flights continued.
  • The integrator clip never acts. In every set, clip arm and horizon the median share of alive steps with the roll/pitch attitude integrator at ±0.3 is 0.000, and removing it changes no divergence count at all (bolted-on: 84 and 84 of 384 at 6 s; 76 and 76 at 18 s). Milestone 7c's explanation is wrong. Removing the ±1 feed-forward clip too (it is active on 36–44% of the bolted-on designs' steps) moves the totals by at most 3 episodes either way.
  • More time does not explain it either. The linear model, iterated from each flight's initial offset alone, reaches 5 m or 60° within 6 s in 60% of the 1.25× designs' unstable flights and within 18 s in 87%; the simulated flights at 1.25× diverge on 0 of 384 at 6 s and 0 of 384 at 18 s, and their median position error falls from 0.110 to 0.084 m.
  • The saturation that is active is the tilt command. At the default 20° cap it is clipped on 72–75% of the bolted-on designs' alive steps and 13–24% of the 1.25× designs', against 0.3–2.7% for the 0.8× and retrained designs.
  • Post hoc (experiments/m7d_clip/diag.py, written after results.txt was read): widening the tilt-command clip releases the instability. Lever 0, 384 flights per set, clip 20 / 30 / 45 / 59°: bolted-on 84 / 360 / 384 / 384 diverged at 6 s; 1.25× 0 / 4 / 166 / 212 at 6 s and 0 / 6 / 318 / 339 (88%, against the 95% the linear model calls unstable) at 18 s. The 0.8× designs: 0 at every clip and horizon. The retrained designs, which sit on the ceiling (0.88–1.11×): 0 / 2 / 13 / 13 at 6 s, 7 of the 13 in the seed at 0.88×. With the 20° clip the 1.25× designs hold position but their tilt tracking error is 8.7–10.0° RMS against 2.6–3.6° at 0.8×: a bounded oscillation against the tilt limit, not a stable loop.
  • Verdicts. V1 held (exact at both levers). V2 refuted (84 → 84). V3 refuted (0 of 367 linear-unstable flights diverged). V4 held, but only because nothing diverges with any clip arm; the rival it tested is refuted by V7. V5 held (0 in every arm). V6 refuted (integrator at its clip on 0.000 of steps in every set). V7 split: within 18 s held (0.87), within 6 s refuted (0.60 against ≤ 0.10). V8 refuted (bolted-on, 18 s, lever 0: 73 with both clips removed against 76).
  • Reading: milestone 7c's "bounded" slow instability was held by the 20° tilt-command limit, the outermost saturation in the position loop, not by the integrator clip and not by the short episode. The linear test's false alarms are real instabilities the tilt limit happens to contain as an oscillation; raise the limit (an aggressive position mode, a racing profile) and they diverge at close to the predicted rate. So the linear ki/kp ceiling is a better rule than 7c gave it credit for, and the 0.8× margin is the only setting tested that survives every tilt limit. For a builder: designs past the ceiling fly only because the tilt limit is tight, and a design tuned onto the ceiling (as CMA-ES does) loses 3–4% of flights once the limit is lifted.
  • Limits: one geometry (above), lever 0 for the tilt sweep, 6 designs per set, same simulator and assumed plant ranges; the tilt sweep is post hoc and exploratory; no describing-function analysis of the limit cycle yet. Preliminary.
  • Next: a describing-function prediction of the oscillation amplitude under the tilt limit, to turn "bounded" into a number; CMA-ES with the 0.8× ceiling as a constraint, scored at 20° and 45° tilt limits; the paper draft's stability section.

2026-11-13 · Milestone 7c: a linear stability test for the side-force feed-forward, checked against the 36 milestone-7b designs (predictions written and committed before experiments/m7c_stab/run.py was run)

  • Why: milestone 7b found that the integrator feed-forward runs away when bolted onto designs whose attitude integrator is strong, and a post-hoc quasi-steady criterion failed to separate the diverging designs. Its first listed next step: a linearised stability analysis of the full attitude-plus-position loop, to give a design rule that can be checked against those designs.
  • Built: src/codesign_cant/linstab.py. Around hover (zero tilt, rate, torque and yaw; no clip or mixer saturation; noise off) every Euler step of poshold.rollout is linear and the two horizontal axes decouple; gust and wind enter additively, so they do not change stability. One axis is an 11-state discrete map (position, velocity, position integrator, tilt and the one-step-old tilt the controller reads, rate and its old copy, attitude integrator, delivered torque through the randomised lag, and milestone 7's torque estimate and filtered feed-forward). The pitch axis has the side force +lat·τ, the roll axis (in −y) −lat·τ, so each vehicle has one axis where the side force helps the tilt and one where it fights it; both are checked. step copies rollout line by line and matrix builds the map from unit vectors. tests/test_linstab.py checks the linear model against rollout itself (noise, gust and wind off, 0.1 mm offset; one milestone-7 design, four plant/feed-forward cases): position RMS agrees to 1e-3. ki_ceiling bisects the largest multiple of a design's attitude ki that keeps every one of the eight corners of the plant randomisation (torque 0.7/1.0, mass 0.9/1.3, lag 1/2× nominal) and both axes stable.
  • What was looked at before writing these predictions: only the test above (one design, not a stability verdict on any of the 36) and milestone 7b's published per-design divergence counts. No spectral radius of any milestone 6b/7/7b design has been computed.
  • Plan: experiments/m7c_stab/run.py. (a) For milestone 7b's designs (6b aware with feed-forward off, 6b+integ, integ retrained with it on and off; both geometries, 6 seeds: 48 design-conditions) and milestone 7's retrained commanded-torque designs (12), compute the spectral radius at each of the 64 held-out episodes' own plant draw (seed 9999), and re-run the held-out rollouts to get per-episode divergence at lever 0 and 1; classify. (b) ki_ceiling for every design. (c) A new, out-of-sample check of the rule: milestone 6b's 12 force-aware designs with the integrator feed-forward on and ki set to 0.8× and 1.25× their own ceiling (24 new design-conditions, never trained or scored before), held-out rollouts at lever 0 and 1.
  • Predictions: - U1: at lever 0, over all 48 × 64 integrator-mode and feed-forward-off episode pairs, "linearly unstable at this episode's plant" classifies divergence with accuracy ≥ 0.95 and catches ≥ 90% of diverged episodes. - U2: all six milestone 6b above designs with the integrator feed-forward are unstable at the worst corner (ceiling multiple < 1); all six retrained above designs have ceiling multiple > 1; every below design has ceiling multiple > 1 (both sets), and every design with the feed-forward off is stable at all eight corners. - U3: the ceiling, as a ki/kp bound, is lower in above than the retrained designs' own ki/kp by less than a factor of 2 (CMA-ES sat near the edge: retrained ceiling multiple between 1 and 2 in at least 4 of 6). - U4: milestone 7's commanded-torque designs: the two below seeds that diverged on 4 of 64 held-out episodes at lever 1 are linearly unstable at one or more of the 64 held-out plants; the other ten are stable at all 64. - U5 (out of sample): at 0.8× the ceiling, none of the 12 rescaled designs diverges on more than 1 of 64 held-out episodes at lever 0, and the total at lever 1 is at most 4. At 1.25× the ceiling, the observed divergence share at lever 0 is within 0.10 of the share of held-out plants the linear model calls unstable, for at least 10 of 12 designs. - U6: at lever 1 (gust torque on), the divergence share is at least the linear prediction minus 0.02 for every design-condition (gust can add divergences through saturation, rarely remove them).
  • Analysis: confusion counts and per-design shares; no statistical test beyond the counts.

### Outcome (experiments/m7c_stab/results.txt, 84 design-conditions × 64 held-out episodes, about a minute on 4 cores; simulated)

  • Analysis bug found and fixed before any verdict was read off. The first full run called every design marginally unstable (ρ = 1 exactly): in integrator mode milestone 7's torque-estimate state is never updated, so it sits in the map as a held state with eigenvalue 1, and at ki = 0 the unused attitude integrator does the same, which set every ki_ceiling to 0. Both are unobservable, so the rollout-matching test could not see them. Fixed (linstab.step zeroes the unused state; the ceiling's lower bracket is checked at 10⁻³ × ki) with a new test (test_integrator_mode_has_no_held_state); the numbers below are from the fixed code. The model and the predictions did not change.
  • Linear instability is necessary for divergence in integrator mode, not sufficient. Over the 48 milestone-7b design-conditions × 64 episodes at lever 0, no linearly stable episode diverged (0 of 2688) and every diverged one was linearly unstable (84 of 84), but 300 unstable episodes did not diverge: accuracy 0.902. All 300 are milestone 6b's above designs with the integrator feed-forward bolted on, which the model calls unstable at every held-out plant (growth rate 3.9–7.4 per s); 11–44% of their episodes diverge, and the rest hold position badly (median episode RMS 0.16–0.23 m against 0.10).
  • CMA-ES tuned the retrained designs onto the linear boundary. Retrained with the integrator feed-forward (above), their ki/kp is 3.7–4.4 per s and the linear ceiling at the eight plant corners is 3.5–4.6: ceiling multiples 1.04, 1.00 (0.998), 1.01, 1.05, 1.11 and 0.88. Milestone 6b's designs, tuned without the feed-forward, sit at 2.3–3.4× their ceiling. Sixteen training episodes were enough to find the edge, not to keep a margin from it; two of six are slightly past it at a corner plant (never at a held-out draw).
  • Out of sample, the margin is what matters. Milestone 6b's six above designs with ki cut to 0.8× their ceiling: no divergence at lever 0 or 1 and median position RMS 0.10 m, the same as without the feed-forward. At 1.25× the ceiling the model calls 80–100% of held-out plants unstable, yet none diverges and median position RMS is 0.11 m: the instability is slow (growth 0.50–0.66 per s) and stays bounded. (The below designs could not be rescaled: their ceilings are ≥ 4× or undefined, so 12 new design-conditions, not 24.)
  • The commanded-torque feed-forward's divergences are linear, episode by episode. For milestone 7's 12 retrained designs, the model names the same two below seeds that diverge, and at lever 0 it calls 10 of 768 episodes unstable, 7 of which diverge, against 1 diverged among the 758 it calls stable (accuracy 0.995). There a slow instability (0.44–0.66 per s) is enough to diverge, unlike in integrator mode, so the growth rate alone does not set the outcome across feed-forward types.
  • Verdicts. U1 refuted (accuracy 0.902 < 0.95; the sensitivity part held, 1.00). U2 split: all six bolted-on above designs unstable held; retrained above all above the ceiling refuted (0.998 and 0.883); every below design above the ceiling refuted (milestone 6b seed 1 is unstable at one corner even at ki ≈ 0, with the feed-forward off too); every feed-forward-off design stable at all eight corners refuted (4 of 24 have ρ 1.00003–1.00029 at a corner, growth ≤ 0.15 per s). U3 held (ceiling multiples in [1, 2] for 4 of 6; ceiling within a factor 2 of the designs' own ki/kp). U4 held. U5: the 0.8× half held, the 1.25× half refuted (observed 0 against predicted 0.80–1.00, all six). U6 refuted (bolted-on designs: predicted 1.00, observed 0.13–0.45).
  • Post hoc (experiments/m7c_stab/diag.py, written after the results were read). Among the 14 design-conditions the model calls unstable, the growth rate ranks the diverged share: Spearman 0.83 at lever 0 (p 0.0002), 0.82 at lever 1. Within the bolted-on designs, episodes that diverge have higher growth rates at their own plant (median 5.3 against 3.9 per s; AUC 0.80, one-sided Mann–Whitney p 10⁻¹⁷). A plausible reason weak integrator-mode instability stays bounded is the attitude-integrator clip (±0.3), which caps the feed-forward; that is not tested here.
  • Reading: an 11-state linear map, checked against the simulator to 10⁻³ and computed in milliseconds, is a usable necessary test for the side-force feed-forward: no design it calls stable at an episode's plant diverged in that episode under either feed-forward. For a builder adding the integrator feed-forward: compute the ki/kp ceiling at the corners of your plant uncertainty and stay at about 0.8× it; an optimiser left alone will sit on the ceiling. Across these designs the ceiling is 3.5–4.6 per s for the above geometry.
  • Limits: linearised about hover at zero yaw; the clips and the mixer are where divergence actually happens, and they are outside the model; a ceiling only for designs that use the integrator (the below ones barely do); same simulator, same assumed lever and plant ranges as milestones 6–7b; six seeds. No closed-form rule in lat and the motor lag yet: the ceiling is numerical, per design. Preliminary.
  • Next: a describing-function or saturation-aware analysis of the integrator clip to explain bounded versus divergent instability; train with the linear ceiling as a constraint (0.8×) and compare against unconstrained CMA-ES on held-out divergence and J; a closed-form approximation of the ceiling in lat, lag and the position gains; the paper draft, where this gives the stability section.

2026-11-13 · Milestone 7b: feed forward only the integrator's torque, so no fast loop closes (predictions written and committed before the full experiments/m7b_integ/run.py run)

  • Why: milestone 7's commanded-torque feed-forward evened the two horizontal axes but fed the proportional and derivative torque back into the tilt reference, closing a loop with the motor lag inside it; under lag and torque uncertainty that added a divergence tail (below: two of six seeds diverged on 4 of 64 held-out episodes each, J −0.33 against milestone 6b). Its first listed next step: feed forward only the steady part.
  • Method (poshold.rollout(ff_mode="integ")): the side force that hurts one axis comes from the steady counter-torque to the gust, and the attitude PID already holds an estimate of that torque in its integrator. The position loop subtracts lat · ki · integ · τ_max,nom / m_nominal (previous step, unfiltered, gain 1) from the acceleration the tilt must make. The P and D torque, which follow the tilt reference within the motor lag, are left out. In steady state the attitude error returns to zero whatever the reference shift, so the integrator is pinned by torque balance and the added loop can only act through the slow integrator. Same numbers as milestone 7 (geometry, the controller's own state), no filter constant: 12 coordinates as in milestones 6/6b.
  • Pre-check: none on held-out data. A smoke run (64 evaluations, 2 seeds, 8 episodes of a different test seed, 4242; not evidence) checked the pipeline, and its output was deleted.
  • Plan: experiments/m7b_integ/run.py. Retrain the force-aware arm with the integrator feed-forward on under both rigid geometries (above, below), milestone 6/6b/7's CMA-ES, budget (3200), 16 training episodes, 6 seeds, two starts: 24 runs. Score on the 64 held-out episodes (seed 9999) at lever 0, 1 and 2, side force on and off; also with the feed-forward off (ablation); milestone 6b's force-aware designs with the integrator feed-forward added (no retraining); milestone 7's retrained designs with their commanded-torque feed-forward, re-scored.
  • Predictions (lever 1 unless stated): - T1 (6b designs + integrator feed-forward, not retrained): late-window hurt/helped axis ratio at most 1.3 (mean) in both geometries (against 1.91 and 3.41 without), and no more diverged episodes in total than the same designs without it. - T2 (H3, retrained): mean on/off position ratio at most 1.05 in both geometries (milestone 7: 1.13 and 1.60). - T3 (the point of the change): retrained with the integrator feed-forward, no seed diverges on more than 1 of 64 held-out episodes in either geometry, and the total diverged episodes are at most milestone 7's in each geometry (milestone 7: 1 above, 8 below). - T4: retrained against milestone 7's retrained designs: mean J better in both geometries. - T5: retrained against milestone 6b's force-aware designs: J better in at least 4 of 6 seeds and whole-episode position RMS lower in at least 4 of 6, both geometries. - T6: retrained against no cant (milestone 6b's nocant): yaw RMS at least 3° lower in 6 of 6 seeds and J better in at least 5 of 6, both geometries. - T7: chosen cant median 30–45° in both geometries. - T8 (lever 2, out of distribution): retrained designs diverge on a smaller mean share of episodes than milestone 6b's force-aware designs, in both geometries (milestone 7 failed this below: 13.0% against 8.6%). - T9: the late-window hurt/helped ratio of the retrained designs is at most 1.3 in both geometries, and switching their feed-forward off raises it above 1.5 in both (the controller term, not the airframe, evens the axes).
  • Analysis: per-arm means and per-seed values, paired differences by seed with sign counts; six seeds, no test beyond the sign count.

### Outcome (experiments/m7b_integ/results.txt, 24 CMA-ES runs of 3200 evaluations, about 50 minutes on 4 cores; simulated)

  • Retrained with it, centre of pressure above: the first side-force fix that costs nothing. No held-out episode diverges at lever 0 or 1 (milestone 6b and 7: one each). Hurt/helped late-window ratio 0.83 (slightly over-cancelled; 2.06 with the feed-forward switched off on the same designs). Mean J −1.58 against −1.68 (6b) and −1.67 (milestone 7), though by sign only 2 and 1 of 6 seeds are better: the mean gain is the one diverging seed the others had. At lever 2 it diverges on 15 episodes against 23 (6b) and 19 (milestone 7), J better than 6b in 6 of 6.
  • Centre of pressure below: it barely acts. Ratio 2.04 against 3.01 with it off and 3.41 for 6b; J −1.72 (6b −1.73). The below designs barely use the attitude integrator (ki/kp 0.05–0.28 per s against 3.7–12.6 above, diag.txt), so the integrator holds little of the steady counter-torque and an integrator-only feed-forward sees little of the side force. It is still safe there: no more divergences than 6b at lever 0 and 1 (2 episodes, the same seed-1 design family), and far fewer than milestone 7 (2 against 8; J −1.72 against −2.05).
  • Added to designs not trained for it, it is dangerous where the integrator is strong. Milestone 6b's above designs (ki/kp 9–13 per s) diverge on 7 to 29 of 64 held-out episodes each at lever 0 (no gust torque at all), 90 of 384 at lever 1; the below designs are untouched. So removing the fast P/D path did not remove the loop: the integrator path "tilt reference → attitude error → integrator → predicted side force → tilt reference" also runs away, slowly, when the integrator gain is high. A post-hoc quasi-steady criterion (runaway when lat·τ_max·kp/(m g) > 1) does not separate the diverging designs (0.68–0.97) from the safe retrained ones (0.94–1.02) (diag.py, written after the results were read); the retrained designs cut ki/kp to 3.7–4.4 per s, and CMA-ES found that on 16 training episodes.
  • T1 refuted in both (above: ratio 1.05, but 90 diverged episodes against 1; below: ratio 2.54). T2 held above (on/off 1.010), refuted below (1.249; median 1.02, one seed). T3 held above (0 diverged), half-held below (total 2 ≤ 8, but seed 1 diverges on 2). T4 held on means in both (J +0.09 and +0.33 against milestone 7), with weak sign counts (1 and 2 of 6). T5 refuted above (J better in 2 of 6, position lower in 2), held below (5 and 4 of 6). T6 held (yaw −5.2° and −7.7° against no cant, J better 6 of 6, both). T7 held (cant median 33.3° and 31.7°). T8 held above (3.9% against 6.0%), refuted below (10.9% against 8.6%). T9 held above (0.83 on, 2.06 off), refuted below (2.04).
  • Reading: whether a feed-forward from the attitude integrator can cancel the canted rotor's side force depends on how much of the steady counter-torque the integrator carries. Where it carries most (the above designs), retraining with it on gives the only variant so far that evens the axes with no divergence at the design wind and fewer out of distribution; bolted onto a design tuned without it, the same term runs away. Where the integrator carries little (below), the term is inert and harmless. For a builder: add it only together with a retune (or an explicit ki/kp ceiling), never as a patch on an existing tune. Cant still pays in every seed (yaw 5–8° lower).
  • Limits: as milestone 7 (assumed lever, one wind state, perfect altitude hold, six seeds, two starts, gain 1, no hardware). The ki/kp explanation is post hoc and correlational (two design families); no stability analysis of the full loop was done. Preliminary.
  • Next: a linearised stability analysis of the full attitude-plus-position loop with the integrator feed-forward, to give a design rule (ki/kp ceiling as a function of lat and the motor lag) that can be checked against these 36 designs; train with a divergence penalty or more episodes; a cant pattern whose side-force map is a rotation; motor data (milestone 3); the paper draft.

2026-11-12 · Milestone 7: can the controller cancel the side force with numbers it already has? A position-loop feed-forward (predictions written and committed before the full experiments/m7_ff/run.py run)

  • Why: milestone 6b found that on a rigid frame the attitude loop's steady counter-torque to the gust makes a side force (lat·τ) that adds to the wind push on one horizontal axis: that axis holds position 1.9–3.4× worse after a wind step, and one seed in six diverges on 1–2 of 64 episodes. Its first listed next step was a feed-forward that cancels it.
  • Method (poshold.rollout(ff=...)): the mixer cannot cancel the force (four rotors, four commands; the side force is fixed by the torque), but the position loop can lean against it. The controller knows lat (its geometry), what the mixer commanded and the nominal motor lag, so it predicts the delivered torque through the nominal lag (as INDI does in aero), turns it into the side force, low-passes it with time constant τ_ff and subtracts F/m_nominal from the acceleration the tilt must make. Causal (previous step's command), nominal plant only (the randomised mass, lag and torque scale stay unknown). No new sensor.
  • Pre-check, seen before these predictions were written (and so not counted as tests): milestone 6b's force-aware designs with the feed-forward switched on, not retrained, scored on the 64 held-out episodes. The axes even out (hurt ≈ helped in nearly every design), but a fast filter destabilises them: τ_ff 0.05 s diverges on 2–70% of episodes, 0.2 s on up to 52%, 0.5 s on 0–17%. The loop "tilt reference → torque → predicted side force → tilt reference" has gain lat·k_p·I/(m g) and the motor lag in it. So the full run lets CMA-ES choose τ_ff as a 13th coordinate (log-uniform 0.05–2 s) instead of us picking it from held-out data. A smoke run (64 evaluations, 2 seeds, 8 test episodes; not evidence) checked the pipeline.
  • Plan: experiments/m7_ff/run.py. Retrain the force-aware arm with the feed-forward on under both rigid geometries (above, below), milestone 6/6b's CMA-ES, budget (3200), 16 training episodes, 6 seeds, two starts: 24 runs. Score on the 64 held-out episodes at lever 0, 1 and 2, force on and off; also with the feed-forward switched off (ablation), and milestone 6b's designs with the feed-forward added at τ_ff 0.2, 0.5, 1 and 2 s (no retraining).
  • Predictions (lever 1 unless stated): - S1 (re-score, 6b designs + feed-forward at τ_ff 1 s): late-window hurt/helped axis ratio between 0.8 and 1.25 (mean) in both geometries, against 1.91 and 3.41 without it; the hurt-axis late RMS falls in at least 5 of 6 seeds per geometry. - S2 (H3, retrained with the feed-forward): mean on/off position ratio at most 1.05 in both geometries. - S3: retrained with the feed-forward against milestone 6b's force-aware designs: whole-episode position RMS lower in at least 4 of 6 seeds per geometry, and J better in at least 4 of 6. - S4: retrained with the feed-forward against no cant (milestone 6b's nocant): yaw RMS at least 3° lower in 6 of 6 seeds and J better in at least 5 of 6, both geometries. - S5: chosen τ_ff median at least 0.3 s in both geometries. - S6: chosen cant median 30–45° in both geometries (the feed-forward removes a cost of cant, so it should not fall below milestone 6b's 31–33°). - S7 (lever 2, out of distribution): the retrained feed-forward designs diverge on fewer episodes (mean share) than milestone 6b's force-aware designs, in both geometries.
  • Analysis: per-arm means and per-seed values, paired differences by seed with sign counts; six seeds, no test beyond the sign count.

### Outcome (experiments/m7_ff/results.txt, 24 CMA-ES runs of 3200 evaluations, about 40 minutes on 4 cores; simulated)

  • The feed-forward evens the axes. Retrained with it, the late-window hurt/helped axis ratio is 1.08 (above) and 1.10 (below), against 1.91 and 3.41 for milestone 6b's designs. Switching the feed-forward off on the same retrained designs brings the split straight back (2.10 and 3.33), so the controller term, not a different airframe, does it. Median hurt-axis late RMS falls from 0.047 to 0.037 m (above) and from 0.040 to 0.030 m (below).
  • It does not pay overall: it trades a steady, one-axis loss for a rare divergence. Above, retrained with it against milestone 6b: J +0.009 (better in 5 of 6), position −0.001 m (lower in 5 of 6), the same single diverged episode. Below: J −0.33 (better in 4 of 6), position +0.036 m (lower in 3 of 6), and two seeds diverge on 4 of 64 held-out episodes each (milestone 6b: one seed, 2). The same retrained below designs with the feed-forward switched off diverge on none and score J −1.52 against −2.05 with it on. Sixteen training episodes did not show CMA-ES the tail the feed-forward creates.
  • Why: the feed-forward closes a loop (tilt reference → attitude torque → predicted side force → tilt reference) with the motor lag inside it, gain lat·k_p·I/(m g), and an uncertain plant (lag 1–2× nominal, torque 0.7–1.0×). A fast filter destabilises it (milestone 6b designs + feed-forward: 16% of episodes diverge at τ_ff 0.2 s below, 3% at 0.5 s, 1.3% at 1 s, 0.8% at 2 s; 0.5% without). A slow filter is safe but only half-cancels the 3 s wind step's side force.
  • S1 held: milestone 6b designs + feed-forward at τ_ff 1 s: hurt/helped 1.22 (above) and 1.20 (below); hurt-axis late RMS lower in 5 of 6 seeds in each.
  • S2 refuted: mean on/off position ratio 1.13 (above) and 1.60 (below), against the 1.05 predicted. Medians 1.01 and 1.02: as in milestone 6b, the typical design does not notice the side force and the mean is the divergence tail.
  • S3 held above, refuted below (position lower in 3 of 6, not 4).
  • S4 held: against no cant, yaw −5.10° (above) and −7.19° (below), 6 of 6 seeds each; J better in 5 of 6 in each. Cant still pays with the feed-forward.
  • S5 held, barely below: chosen τ_ff median 0.75 s (above) and 0.31 s (below; 0.17–2.0 s). The below designs that chose fast filters (0.17, 0.18 s) are the two that diverge.
  • S6 held: chosen cant median 34.4° (above) and 32.9° (below), 1–2° more than without the feed-forward.
  • S7 held above, refuted below: at lever 2 the retrained feed-forward designs diverge on 4.9% of episodes against 6.0% (above) but 13.0% against 8.6% (below).
  • Four of seven predictions held in both geometries (S1, S4, S5, S6); S2 was refuted in both, S3 and S7 in one.
  • Reading: the side force of a canted rotor can be cancelled with numbers every flight controller has, and doing so removes the one-axis position penalty in the typical design. But the cancellation term feeds the commanded torque back into the attitude reference, which adds a divergence tail under motor-lag and torque uncertainty that a 16-episode training set did not expose. For a builder: if you add it, filter it slowly (1 s or more); the gain is millimetres to a centimetre of late-window error on one axis.
  • Limits: as milestone 6b (assumed lever, one wind state, perfect altitude hold, six seeds, two starts, no hardware), plus a feed-forward of fixed gain 1 with only its filter tuned. Preliminary.
  • Next: feed forward only the steady part, the attitude integrator's state (which is the controller's estimate of the gust torque), instead of the whole commanded torque, so no fast loop closes; train on more episodes, or penalise divergence, so CMA-ES sees the tail; a cant pattern whose side-force map is a rotation; RotorPy (milestone 5); motor data (milestone 3).

2026-11-12 · Milestone 5: does the design ranking survive RotorPy with canted rotors? (H5; predictions written and committed before experiments/m5_rotorpy/run.py was run)

  • Why: every result so far comes from one stylised simulator (Euler-angle rigid body, a first-order torque lag, a gust torque rather than aerodynamics, perfect altitude hold). H5 asks whether the co-designed ranking survives a higher-fidelity, independently written simulator. Milestone 3 (motor data) stays blocked on sources, so this run takes milestone 5.
  • Built: src/codesign_cant/rotorpy_twin.py. RotorPy 3.0.0 (Folk et al. 2023, MIT, pip install rotorpy==3.0.0) assumes every thrust axis is body z. CantedMultirotor subclasses its Multirotor and replaces only the wrench: each rotor thrusts along its canted axis n_i from allocation.layout, its rotor drag (H force) and induced inflow act in the rotor's own frame, its reaction torque turns with n_i, and the frame drag can act at a centre of pressure above the centre of mass. Everything else is RotorPy's: quaternion rigid body with gyroscopic terms, first-order motor speed dynamics with thrust ∝ ω², solve_ivp (RK45). Tests check that at α = 0 the wrench equals RotorPy's own compute_body_wrench to 1e-12 (aero on and off), that hover thrust balances weight with zero torque and side force at 0–60°, and that the mixer's thrust change delivers the requested torque.
  • Parameters: mass, inertia, arm, propeller, hover and maximum thrust, motor lag (as the motor time constant) and frame area from aero.size, the same as the stylised model. The aerodynamic coefficients RotorPy needs and the stylised model lacks come from RotorPy's AscTec Hummingbird file, scaled to each propeller by momentum-theory exponents: k_η (C_T = 0.1145, implied by the Hummingbird's k_η at D = 0.2 m) with D⁴, rotor drag k_d and inflow k_z with D³, translational lift k_h with D². The scaling is an assumption; one condition doubles the coefficients.
  • Controller: the stylised laws re-implemented for one vehicle with the same gains, the same one-step-old noisy measurements, the same attitude-priority mixer and the same INDI nominal model, plus an altitude loop (2 rad/s, critically damped; assumed) because RotorPy does not hold altitude for free.
  • Plan: experiments/m5_rotorpy/run.py. The 60 milestone-4 designs (PID, INDI, geometric × cant bound {0, 80°} × 10 seeds), unchanged and not retuned, on the same held-out task (32 episodes of seed 9999; the stylised scores are recomputed at 32). Four RotorPy conditions: torque (aero off, the stylised gust torque injected: isolates rigid-body, motor and altitude effects), aero (RotorPy aerodynamics driven by the stylised wind as a world wind, frame drag at the centre of mass), aero_cop (frame drag at the stylised lever 0.05 + 0.3 L above), aero_cop2 (rotor-drag coefficients doubled). Score J = −RMS − 0.3 P/100 W as in milestone 4.
  • A smoke run (6 designs, seed 0, 4 episodes; not evidence) checked the pipeline: RotorPy RMS 0.4–1.2° above stylised, cant advantage 4.2–5.2° in every condition, no divergences. The predictions below were written after seeing it.
  • Predictions: - F1 (H5): Spearman ρ of J over the top ten (by stylised J) is at least 0.5 in torque and in at least two of the three aero conditions. Risk: the top ten are likely all canted designs within about 1° of each other, so a shift of that size could scramble them. - F2: ρ of J over all 60 designs is at least 0.85 in every condition. - F3: in torque, mean RMS is 0.3–1.5° above the stylised mean (motor-speed dynamics, a real altitude loop). - F4 (H2/H4 carry over): for each law and condition, cant lowers RMS in at least 9 of 10 seeds, and the mean advantage is within 1° of the stylised one. - F5 (H4's refutation test): INDI without cant is worse than PID with cant in at least 9 of 10 seeds in every condition. - F6: divergent episodes (|roll| or |pitch| ≥ 60°) are under 1% of episodes in every condition.
  • Analysis: Spearman ρ (scipy), per-arm means, paired differences by seed with sign counts. No test beyond the counts.

### Outcome (experiments/m5_rotorpy/results.txt, 60 designs × 4 conditions × 32 episodes, about 60 minutes on 4 cores; simulated)

  • H5 is refuted as written, and the post-hoc check shows the criterion cannot be met. Spearman ρ of J over the top ten (by stylised J) is 0.21 (torque), 0.20 (aero), 0.30 (aero_cop) and 0.25 (aero_cop2), all below 0.5. F1 refuted. But the top ten are all canted INDI or geometric designs whose stylised J spans 0.064 (−13.35 to −13.29), so they are tied. Re-scoring them in the stylised simulator on five fresh 32-episode sets gives a top-ten ρ of 0.26 on average (−0.22 to 0.47; noise_ceiling.py, written after reading the outcome). RotorPy agrees with the stylised model about as well as the stylised model agrees with itself on new episodes. The top-ten test measures episode noise, not fidelity.
  • The ranking that a designer acts on survives. Over all 60 designs ρ is 0.80 (torque) and 0.88–0.91 (aero conditions), against a self-agreement ceiling of 0.95–0.97. F2 held in the aero conditions and failed in torque (0.799). The six law × cant arms keep their order: ρ 1.00 (aero, aero_cop2) and 0.94 (torque, aero_cop). Geometric control with cant is the best arm in every simulator.
  • Cant still wins in RotorPy, in 10 of 10 seeds for every law in every aero condition: PID −4.8 to −5.8° (stylised −5.6°), geometric −4.7 to −5.1° (−4.9°), INDI −3.1 to −3.5° (−4.2°). In torque it wins 10 of 10 for PID (−7.3°) and geometric (−6.0°) but only 8 of 10 for INDI (−1.9°). F4 held in the aero conditions, except INDI under aero (1.03° from stylised against a 1° margin), and failed in torque.
  • INDI loses most in the move. Canted INDI designs lose 1.8–2.1° of RMS in the aero conditions and 4.2° in torque, against 1.4–1.8° and 2.9° for PID and 0.9–1.1° and 2.8° for geometric control. INDI's increment rests on its model of the actuator: a first-order torque lag. In RotorPy the actuator is a first-order speed lag with thrust ∝ ω². The model-based law is the one most hurt when the model changes, and its paper claims need the high-fidelity check.
  • The stylised gust torque is harsher than RotorPy's aerodynamics. torque (stylised disturbance, RotorPy body and motors) adds 3.4° of RMS on average and 54 divergent episodes of 1920 (2.8%), 43 of them on uncanted designs. F3 (0.3–1.5° added) and F6 (under 1%) failed there. With RotorPy's own aerodynamics driven by the same wind, the added error is 1.3–1.5° and there are 1–3 divergent episodes (≤ 0.16%); F6 held. Two reasons, neither tested separately: in RotorPy the vehicle drifts with the wind, so airspeed and drag fall, while the injected torque does not; and the randomised thrust loss (0.7–1.0) makes the altitude loop raise the collective, eating torque headroom the stylised model kept.
  • F5 held everywhere: INDI without cant is worse than PID with cant in 10 of 10 seeds in every condition (+3.1 to +3.9°). H4 survives the move.
  • Doubling the rotor-drag coefficients (aero_cop2) changes mean RMS by under 0.25° against aero_cop: the conclusions do not hang on the assumed drag scaling.
  • Reading: the stylised simulator ranks design classes correctly (cant over no cant, geometric ≥ INDI ≈ PID with cant) and individual designs well (ρ 0.88–0.91 against a ceiling of 0.95–0.97). It cannot rank designs that it scores within 0.1 of each other, and neither can it against itself. It is optimistic by 1–2° of RMS, most for INDI.
  • Proposed (post hoc, for the paper): replace H5's top-ten test with (a) arm-level rank agreement and (b) all-design ρ as a fraction of the simulator's own fresh-episode ceiling, with a pre-registered threshold for the next fidelity step.
  • Limits: designs are not retuned in RotorPy (a retune would likely recover part of INDI's loss); the aerodynamic coefficients are the Hummingbird's scaled by momentum-theory exponents, not measured for these propellers; the altitude-loop gains are assumed; 32 episodes; one RotorPy version (3.0.0) with RK45; no hardware. Preliminary.
  • Next: retune the top designs inside RotorPy to separate model mismatch from overfitting; give INDI a speed-lag actuator model; motor data (milestone 3); paper draft (milestone 7).

2026-11-12 · Milestone 6b: which gust-torque sign does a real frame have? Neither: a rigid frame is helped on one axis and hurt on the other (predictions written and committed before experiments/m6b_cop/run.py was run)

  • Why: milestone 6 found that H3 hangs on the sign of the gust torque (on/off position ratio 0.96 with the inherited sign, 1.21 flipped) and left "derive which sign a real frame has" as the next step.
  • Derivation (poshold.cop_signs, checked in tests/test_poshold.py against np.cross): drag F = (F_x, F_y, 0) acting at the centre of pressure r = (0, 0, h), z up, gives τ = r × F = (−h F_y, h F_x, 0). In rollout the roll torque goes with F_y and the pitch torque with F_x, each with its own sign, so a rigid body has signs (roll, pitch) = (−sign h, +sign h). Milestone 6's (+1, +1) and its flip (−1, −1) are not rigid bodies: each puts the centre of pressure below the centre of mass on one axis and above it on the other. The side-force map from the allocation rows is F_y = lat·τ_x, F_x = lat·τ_y (a reflection, not a rotation, because the canted layout is chiral), so the steady counter-torque −τ makes a side force −lat·(ds_roll·k)F_y on y and −lat·(ds_pitch·k)F_x on x, k = lever. With opposite signs, the side force opposes the drag on one axis and adds to it on the other for either sign of h (h above: y hurt, x helped; h below: x hurt, y helped). The magnitude of the effect is lat·lever, about 0.6–0.8 of the drag at milestone 6's designs. So the question "which sign" becomes "how large is the lever", and the hurt axis should show it.
  • Plan: experiments/m6b_cop/run.py. Retrain milestone 6's aware and nocant arms under both rigid geometries (above, below), same 12-parameter CMA-ES, budget (3200), training episodes (16), 6 seeds and two starts: 48 CMA-ES runs. Score each design on the 64 held-out episodes with the side force on and off, at lever 0, 0.5, 1 and 2 times the inherited 0.05 + 0.3 L (assumed; no frame data), with late-window (3–6 s) position RMS per world axis. Also re-score milestone 6's designs under both geometries (no retraining).
  • A smoke run (64 evaluations, 2 seeds, 4 training and 8 test episodes; not evidence) checked the pipeline. It showed milestone 6's designs with a late-window hurt/helped axis ratio of about 3 under below. The predictions below were written after seeing it.
  • Predictions (all at lever 1 unless stated): - R1 (re-score, milestone 6 aware designs): the on/off position ratio lies between milestone 6's two values (0.96 and 1.21) in both geometries; the late-window hurt/helped axis ratio is at least 1.5. - R2 (H3, retrained aware): mean on/off ratio at most 1.2 in both geometries, so H3 is not refuted for a rigid frame at the inherited lever. - R3: retraining under the geometry lowers the on/off ratio against the re-scored milestone 6 design in at least 4 of 6 seeds, in each geometry. - R4: cant lowers held-out yaw RMS by at least 3° in 6 of 6 seeds and improves J in at least 5 of 6, in both geometries. - R5 (lever sweep, retrained aware): the on/off ratio at lever 2 exceeds that at lever 1 in at least 5 of 6 seeds, and its mean crosses 1.2 in at least one geometry. - R6: late-window hurt/helped axis ratio within 0.85–1.15 for nocant (no side force, so nothing breaks the symmetry) and above 1.3 for aware. - R7: chosen cant median 25–45° in both geometries.
  • Analysis: per-arm means and per-seed values, paired differences by seed with sign counts; six seeds, no test beyond the sign count.

### Outcome (experiments/m6b_cop/results.txt, 48 CMA-ES runs of 3200 evaluations, about 75 minutes on 4 cores; simulated)

  • The axis split is real and large. For retrained force-aware designs at the inherited lever, the late-window (3–6 s) position RMS on the axis where the side force adds to the drag is 1.91× (above) and 3.41× (below) the other axis. Without cant the two axes match (0.96 in both). With the force switched off the hurt axis drops from 0.054 to 0.030 m (above) and from 0.057 to 0.021 m (below). R6 held.
  • Without the gust torque the side force costs nothing. At lever 0 the on/off ratio is 1.002 in both geometries, all 12 seed-designs at 1.00. The whole cost of the side force runs through the attitude loop's steady counter-torque to the gust.
  • R1 held: milestone 6's designs re-scored under the rigid geometries give on/off 1.208 (above; per seed 1.04–1.69) and 1.071 (below), between milestone 6's 0.96 and 1.21. Hurt/helped axis ratio 2.08 and 2.76.
  • R2 held above, refuted below, by one seed each way. Mean on/off is 1.142 (above) and 1.261 (below), but the medians are 1.028 and 1.033. Five of six seeds sit at 1.02–1.04 in each geometry. The mean is carried by one seed per geometry (1.72 above, 2.41 below), where 1 or 2 of the 64 held-out episodes diverged and were scored at the 5 m cap. H3's 20% line is crossed by the mean below, not by the typical design: the risk is a rare divergence, not a steady loss.
  • R3 held: retraining under the geometry lowers the on/off ratio against the re-scored milestone 6 design in 5 of 6 seeds in each geometry (the outlier seed goes the other way below, so the mean difference is +0.19 there).
  • R4 held: cant lowers yaw RMS by 5.10° (above) and 7.73° (below), 6 of 6 seeds each, and improves J in 5 of 6 (above, +0.26) and 6 of 6 (below, +0.66). Cant still pays for a rigid frame.
  • R5 held, with a caveat: at twice the lever the on/off ratio rises in 5 of 6 seeds in each geometry, and its mean is 1.36 (above) and 1.19 (below). But lever 2 doubles the gust torque on designs trained at lever 1: 6–27% of episodes diverge, including 10–27% without cant, so lever 2 is mostly an out-of-distribution attitude stress test. There cant's J advantage grows (+1.23 above, +6.36 below), because the uncanted vehicle diverges more.
  • R7 held: chosen cant median 33.2° (above) and 31.1° (below), 3–6° less than under milestone 6's non-rigid signs (36–38°).
  • All seven predictions held except R2 in the below geometry, which failed on one seed's divergences.
  • Reading: no frame has milestone 6's helpful sign on both axes. A rigid canted quadrotor with its centre of pressure off the centre of mass holds position about 2–3× worse along one horizontal axis after a wind step, whichever side the centre of pressure is on. The averaged cost stays small for typical designs (median 1.03×). The tail is worse: one seed in six diverges on 1–2 episodes in 64. The axis is set by the chirality of the cant pattern together with the side of the offset. A designer can choose which axis is hurt but cannot remove the effect except by shrinking the lever (centre of pressure at the centre of mass) or the cant.
  • Limits: the lever and its scale are assumed (0.05 + 0.3 L, swept 0–2×), not measured for a real frame; one wind state drives drag and torque; the disturbance torque is applied from world-frame wind to body axes (yaw ±30° mixes the axes partly); perfect altitude hold; six seeds and two starts; lever 2 was not retrained; no hardware. Preliminary.
  • Next: a feed-forward that cancels the counter-torque's side force (the mixer knows lat), or a cant pattern whose side-force map is a rotation rather than a reflection; measure a real frame's centre of pressure (wind-tunnel or CFD literature) to pin the lever; RotorPy check (milestone 5); motor data (milestone 3).

2026-11-12 · Milestone 6: does the side force of a canted rotor cost position hold more than cant's yaw gain is worth? (predictions written and committed before experiments/m6_poshold/run.py was run)

  • Why: H3. A canted rotor makes a horizontal body force whenever the mixer makes roll or pitch torque. codesign-aero-cant-poshold tested this with a per-axis torque box, a side force √2 too large (tan α / (L/√2)) and 2 s episodes, too short for drift. Milestone 3 (motor data) stays blocked on sources (see milestone 4's entry), so this run takes milestone 6.
  • Model (src/codesign_cant/poshold.py): the side force is derived from the allocation rows, F_y = lat·τ_x and F_x = lat·τ_y with lat = sin α / (L cos α − κ sin α), and yaw torque makes none (tests check it against allocation.layout on 100 random designs). It acts on the delivered torque, after the attitude-priority mixer (collective at hover) and the motor lag. Episodes are 6 s: start 0.5 m off, yaw steps ±30° every 1.5 s, and the mean wind steps by ±3 m/s per axis at 3 s. Translation, drag and noise are -poshold's (assumed, inherited). J = −sqrt((pos/0.10 m)² + (yaw/10°)²) − 0.03·P_hover/100 W, which is milestone 4's 0.3° per 100 W in composite units.
  • Plan: experiments/m6_poshold/run.py, 12-parameter CMA-ES (design, six attitude PID gains, three position gains), 3200 evaluations, 16 training episodes, 6 seeds, two starts per seed (cant at 0.4 and 0.8 of the bound), arms nocant (α = 0), aware (α ≤ 80°, side force in training) and unaware (α ≤ 80°, side force off in training). Held-out: 64 episodes (seed 9999), each design scored with the force on, off, and in calm air (no wind step).
  • A smoke run (32 evaluations, 2 seeds, 4 episodes; not evidence) hinted that the force may help the force-aware designs (position RMS on/off 0.80). The predictions below were written after seeing it.
  • Predictions: - Q1 (H3): for aware designs, held-out position RMS with the force on is at most 1.2× the same designs with it off (mean ratio over seeds). H3 is refuted if the mean ratio exceeds 1.2. - Q2: unaware designs scored with the force on have higher position RMS than aware designs in at least 5 of 6 seeds. - Q3: cant lowers held-out yaw RMS against no cant by at least 3°, in at least 5 of 6 seeds (force on). - Q4: cant improves J (force on) against no cant in at least 4 of 6 seeds. - Q5: cant does not buy position: the paired position RMS difference (aware minus no cant, force on) lies within ±0.05 m. - Q6: the chosen cant under position hold has median 25–45° (milestones 2 and 4: 28–37°). - Q7: the wind step costs position: late-window (3–6 s) position RMS is at least 10% higher than in calm air for every arm (mean over seeds).
  • Analysis: per-arm means and per-seed values; paired differences by seed with SD and sign counts. Six seeds, no test beyond the sign count.

### Outcome (experiments/m6_poshold/results.txt, 36 CMA-ES runs of 3200 evaluations, about 55 minutes on 4 cores; simulated)

  • Q1 held (H3 not refuted in the inherited geometry): for force-aware designs, position RMS with the side force on is 0.961× the same designs with it off (per seed 0.90, 1.24, 0.89, 0.89, 0.94, 0.91); one seed of six crosses 1.2. After the wind step the force helps: 0.714× in the 3–6 s window.
  • Q2 refuted: designs trained without the force are worse than force-aware designs once it is switched on in only 2 of 6 seeds by position (mean +0.034 m, SD 0.137), and lose J in 4 of 6 (mean −0.35, SD 1.28). Two seeds carry the difference (0.387 against 0.097 m and 0.232 against 0.316 m go opposite ways). Ignoring the force is not the disaster -poshold found (1.07 m), partly because the corrected force is √2 smaller.
  • Q3 held: cant lowers yaw RMS by 4.71° (SD 0.96), 6 of 6 seeds (14.88° to 10.18°).
  • Q4 held, barely: cant improves J in 4 of 6 seeds, mean +0.06 (SD 0.84): not distinguishable from zero at six seeds.
  • Q5 held: cant minus no cant in position is +0.034 m (SD 0.093), cant lower in 4 of 6; one aware seed (0.316 m) carries the mean. Cant buys yaw, not position, as in -poshold.
  • Q6 held: chosen cant 36–38° (median 38°); the force-ignoring arm picks 33–37°.
  • Q7 refuted for the canted arms: the wind step raises late-window position RMS above calm air by 22% without cant, but only 9% (aware) and 5% (unaware) with cant. The position integrator absorbs a 3 m/s step within the 3 s left.
  • Found, not predicted: why the force helps. In the inherited model the gust that pushes the vehicle sideways also makes a roll/pitch torque of fixed sign (centre of pressure on one side of the centre of mass). The attitude loop's steady counter-torque then makes a side force that opposes the drag. With lat ≈ 7.8 N per N·m and a centre-of-pressure lever of about 0.1 m, it cancels a large share of the wind push for free. That is a geometric assumption, not a property of cant.

### Post-hoc check, written after reading the outcome (experiments/m6_poshold/dist_sign.py, dist_sign.txt; re-scoring only, no retraining)

  • With the gust torque's sign flipped (centre of pressure on the other side, so the counter-torque's side force adds to the drag), the same force-aware designs lose 1.206× position RMS with the force on (per seed 1.18–1.24, all six at or near the H3 line) and 1.62× in the late window. H3 is refuted in that geometry by the proposal's own 20% criterion, though without retraining for it.
  • Cant still wins the composite there, in 6 of 6 seeds (J +0.19), because the 4.7° yaw gain is worth more than the position it costs. The force-ignoring designs lose 1.59× (up to 2.64×).
  • Reading: whether the side force costs position hold depends on where the centre of pressure sits relative to the centre of mass. In one placement the force fights the wind; in the other it adds to the wind push and costs about 20%. Which physical placement maps to which sign of the inherited model has not been derived here; that is the next step. Either way, cant's yaw gain still pays under the composite. Frame designers can control this sign, and it belongs in the paper.
  • Limits: inherited translation, drag and noise model; the gust torque and drag share one wind state; perfect altitude hold; six seeds and two starts; the composite weights (0.1 m ↔ 10°) and λ are choices; no retraining under the flipped sign; no hardware. Preliminary.
  • Next: retrain under the flipped sign and sweep the centre-of-pressure lever; a RotorPy check of the top designs (milestone 5); motor catalogue data (milestone 3) when a source is reachable.

2026-11-12 · Milestone 4: does a stronger attitude law shrink the cant advantage? INDI and geometric control (predictions written and committed before experiments/m4_controllers/run.py was run)

  • Why: every result so far is for one per-axis PID law. H4 says a stronger law narrows but does not remove the cant advantage, and is refuted if INDI without cant matches PID with cant. A reviewer at RA-L will ask this first: maybe cant only patches a weak controller.
  • Order: milestone 3 (motor catalogue data) was attempted first and deferred. T-Motor's site is reachable but its thrust tables are images and its prices sit on a store that returns 403 through the proxy; the Tyto Robotics database serves its tests through an internal search API that this session may not call. Milestone 4 needs no outside data.
  • Built: aero.rollout(ctrl=...) with two new laws, both reading the same noisy, one-step-old angles and rates as the PID law and sending their torque request through the same mixer. - indi (Smeur, Chu and de Croon 2016): attitude P loop to a rate demand, rate P loop to a desired angular acceleration ν; torque = (filtered estimate of the delivered torque) + g·I_nom·(ν − filtered angular acceleration). The torque estimate runs the command through the design's nominal lag and limits; the plant's randomised torque scale (0.7–1.0), inertia (0.9–1.3×) and lag (1–2×) stay unknown to it. Gyro and torque estimate share a tuned two-pole low-pass filter. Tuned: K_att and K_rate for roll/pitch and for yaw, filter cut-off (5–100 Hz), effectiveness scale g (0.5–2). - geo (Lee, Leok and McClamroch 2010): τ = I_nom(−k_R e_R − k_Ω Ω − k_I ∫e_R) + Ω × I_nom Ω, with the SO(3) error e_R = vee(R_dᵀR − RᵀR_d)/2. Tuned: k_R, k_Ω, k_I for roll/pitch and for yaw.
  • Plan: experiments/m4_controllers/run.py. Law {pid, indi, geo} × cant bound {0, 80°}, attitude-priority mixer with collective at hover, TWR 2, λ = 0.3° per 100 W (milestone 2's middle price), 10 seeds, three CMA-ES starts per seed (cant started at 4, 32, 64°), 1920 evaluations per start, best training score kept: 180 CMA-ES runs. The PID arms use restarts.py's CMA-ES seeds, so their seeds 0–4 must reproduce milestone 2. Held-out: 64 wide-plant episodes, plus the same designs at twice the gust intensity (gust_std 8) as an out-of-distribution check.
  • Predictions: - S1 (reproduction): PID seeds 0–4 give milestone 2's restart numbers (held-out RMS 17.29, 17.45, 19.84, 21.08, 18.48° without cant; 13.17, 13.51, 13.26, 12.94, 12.96° with cant) exactly. - S2: INDI beats PID on held-out RMS in both cant arms, paired, in at least 8 of 10 seeds (it cancels the gust torque and the inertia mismatch that PID must absorb with gain). - S3 (H4, first half): the cant advantage narrows under INDI against PID (paired shrinkage positive in at least 7 of 10 seeds). - S4 (H4, second half): cant still helps under INDI: cant minus no cant negative in at least 8 of 10 seeds, by more than 1° on average. - S5 (H4's refutation test): INDI without cant does not match PID with cant: INDI-no-cant RMS is above PID-cant RMS in at least 8 of 10 seeds. Reason: without cant the vehicle is authority-limited in yaw (milestone 2: yaw scaled on 100% of steps), and no law creates torque the rotors cannot make. - S6: geometric control is within ±1° of PID in both arms (milestone 0 found gyroscopic compensation worth nothing at these angles, and e_R ≈ the Euler error below 30°). - S7: the chosen cant under INDI and geo has a median of 25–40°, as under PID. - S8: at twice the gust intensity the cant advantage keeps its sign for all three laws (mean negative, at least 8 of 10 seeds).
  • Analysis: per-arm means and per-seed values of held-out RMS, roll and yaw RMS, gust-8 RMS, composite J, hover power, chosen cant, mixer saturation shares; paired differences by seed with SD and sign counts. With 10 seeds a sign count of 9 or 10 is p ≤ 0.021 two-sided (sign test); stated where it applies, no other test.

### Outcome (experiments/m4_controllers/results.txt, 180 CMA-ES runs, about 45 minutes on 4 cores; simulated)

  • S1 held: the PID arms reproduce milestone 2's restart numbers exactly for seeds 0–4 (17.29 … 18.48° and 13.17 … 12.96°).
  • S2 held, narrowly with cant: INDI beats PID by 1.25° without cant (SD 1.20, 10 of 10, sign test p = 0.002) and by 0.16° with cant (SD 0.19, 8 of 10).
  • S3 held: under INDI the cant advantage narrows by 1.09° against PID (SD 1.23, 9 of 10).
  • S4 held: cant still helps under INDI by 4.18° (SD 0.25, 10 of 10, p = 0.002); under geometric control by 4.88° (SD 0.20, 10 of 10); under PID by 5.27° (SD 1.32, 10 of 10).
  • S5 held; H4 survives its refutation test: INDI without cant is 4.02° worse than PID with cant (SD 0.31, 10 of 10 seeds); geometric without cant is 4.53° worse. Without cant the mixer still scales yaw on 54% (INDI) to 60% (geo) of held-out steps; no law makes yaw torque the rotors cannot.
  • S6 held: geometric minus PID is −0.74° without cant (SD 1.22, 5 of 10) and −0.35° with cant (SD 0.17, 10 of 10), within ±1°.
  • S7 held: median chosen cant 33.9° (INDI) and 35.3° (geo), against 34.2° (PID). The cant optimum does not depend on the attitude law.
  • S8 held: at twice the gust intensity cant wins in 10 of 10 seeds under every law: −9.82° (PID), −4.51° (INDI), −5.78° (geo).
  • Found, not predicted: most of what INDI buys is consistency in the uncanted vehicle. PID without cant ranges 17.2–21.1° across seeds (SD 1.2); INDI 16.9–17.6° (SD 0.2). At twice the gust intensity INDI beats PID by 5.36° without cant but by 0.05° with cant. So about half of PID's cant advantage under strong gusts (9.8° to 4.5°) is a controller effect that INDI also removes; the other half, and all of it at the design gusts bar 1.1°, is the actuator.
  • Reading: H4 is supported in this simulator. A stronger attitude law narrows the cant advantage (by 1.1° with INDI, 0.4° with geometric control) and does not remove it: cant is worth about 4.2° to 4.9° of held-out RMS whichever law flies the vehicle, and the co-designed cant stays at about 34°.
  • Limits: one power price (0.3° per 100 W), TWR fixed at 2, the attitude-priority mixer only; INDI uses the nominal actuator model with no adaptive effectiveness estimation (Smeur et al.'s adaptive part is not implemented) and the same one-step measurement delay as the other laws; the geometric law has no feed-forward because the references are steps. Mass, cost and motor fits remain inherited assumptions (milestone 3 still open). Simulated.
  • Next: milestone 3 (motor catalogue data) needs a data source this session could not reach: T-Motor's thrust tables are published as images, and its prices are behind a store the proxy blocks. A manual transcription of a dozen datasheets with the source of each row, or a measured open dataset, would unblock it. Otherwise milestone 5 (RotorPy check of the top designs).

2026-11-12 · Milestone 2: combined commands on one thrust polytope, and a price on hover power (predictions written and committed before experiments/m2_polytope/run.py was run)

  • Why: milestone 1 kept the per-axis torque box: roll, pitch and yaw could each saturate at their pure-axis limit at the same time. A real mixer cannot do that. The roll, pitch, yaw and Fz rows of the canted-X wrench map are orthogonal sign patterns (a 4×4 Hadamard matrix), so with the collective held at hover the attainable set is |x ± y ± z|-type: per rotor, −Th ≤ (√2/4kr)(±τr ± τp) + (1/4ky)(±τy) ≤ Tmax − Th. At TWR 2 a full roll command alone uses all the thrust swing and leaves no yaw. Milestone 1 also found the cant optimum held up by an attitude-only objective that leaves hover power free up to the $300 cap.
  • Built: allocation.mix_priority, an attitude-priority mixer (roll and pitch scaled together first, yaw gets what is left, as PX4 and Betaflight do; collective held at hover, or free as in airmode) with exact largest-scale solutions; tests check delivered torque against layout, thrust bounds, hover Fz, and that 0.1% more of either scaled part leaves the polytope. aero.rollout(mixer=...) and aero.score(lam=...), J = −RMS − λ·P_hover/100 W.
  • Plan: experiments/m2_polytope/run.py, the milestone-1 co-design (1920 evaluations, 5 seeds, 64 held-out episodes, c = 0.081, actuator hover, TWR 2) over mixer {box, prio} × cant bound {0, 80°} × λ {0, 0.1, 0.3, 1} deg per 100 W, plus prio_free (collective free) at λ {0, 0.3} × cant {0, 80°}. 100 CMA-ES runs. λ spans the range where the energy term goes from negligible (0.1: a 1.5 kW design pays 1.5°) to dominant (1: 15°); milestone-1 designs drew 160 W to 1.5 kW at near-equal RMS, so hover power was free to the optimiser.
  • Predictions: - Q1 (reproduction): box, λ = 0 reproduces milestone 1's hover/TWR 2 arms (15.44° no cant, 10.63° cant ≤ 80°) exactly. - Q2: under prio the held-out RMS rises against box for both no cant and cant (combined commands lose authority), by more than 1° for no cant. - Q3: cant's paired advantage at λ = 0 is larger under prio than under box (more negative than −4.81°), because the larger ky needs less thrust swing per unit yaw torque, so yaw is starved less by roll and pitch. - Q4 (tests H1): with energy priced the chosen cant falls; at λ = 0.3 the median chosen cant is below 25° under prio (H1 predicts 5–20°). - Q5: hover power falls with λ in both arms; at λ = 0.3 the 5-seed mean hover power is below half its λ = 0 value. - Q6: cant still wins on the composite objective at λ = 0.3 under prio (paired composite better in at least 4 of 5 seeds), and at λ = 1 it no longer wins on RMS by more than 2°. - Q7: prio_free (airmode) beats prio on RMS at λ = 0 for both cant arms.
  • Analysis: per-arm means and per-seed values of RMS, composite score, hover power, mass, chosen cant; paired cant-minus-no-cant differences in RMS and composite by seed with SD and sign count; mean share of steps in which the mixer scaled yaw (s_yaw < 1) on the held-out set. Five seeds, no test beyond the sign count.

### Outcome of run.py (experiments/m2_polytope/results.txt, 100 CMA-ES runs, about 20 minutes on 4 cores; simulated)

  • Q1 held: box, λ = 0 reproduces milestone 1 exactly (15.44° and 10.63°).
  • Q2 held: the shared polytope costs accuracy. prio minus box, paired by seed at λ = 0: +2.90° with no cant (SD 0.29, 5 of 5), +1.63° with cant (SD 0.06, 5 of 5). Without cant the tuned controller saturates roll and pitch on 97% of held-out steps, and yaw gets nothing on 100% of them; with cant roll and pitch saturate on 0.7% and yaw is scaled on 39%.
  • Q3 held: cant's paired advantage at λ = 0 grows from −4.81° (box) to −6.08° (prio, SD 0.28, 5 of 5).
  • Q4 refuted: at λ = 0.3 under prio the chosen cants are 35, 0, 13, 31, 34° (median 31°, not below 25°). Under box cant stays at 29–36° even at λ = 1. The prio split looks like an optimiser failure, not an optimum: the low-cant seeds score 4–6° worse on both training and held-out than the seeds that kept cant (λ = 0.3: J −18.9 and −21.7 against −13.8 to −14.0).
  • Q5 held, and more: hover power falls by 4–7× from λ = 0 to λ = 0.3 in every arm (box cant: 1380 W to 219 W) while RMS moves by at most 0.45° (box). Milestone 1's heavy designs were not buying accuracy with that power; the attitude-only objective was flat in power.
  • Q6 held: at λ = 0.3 under prio cant wins the composite in 4 of 5 seeds (mean −2.65), and at λ = 1 its RMS edge is −0.85° (under 2°) because 4 of 5 seeds fell to low cant.
  • Q7 held (cant 4 of 5): airmode (prio_free) beats prio by 1.56° with no cant (5 of 5) and 0.26° with cant (4 of 5) at λ = 0.
  • Follow-up, specified before running: experiments/m2_polytope/restarts.py reruns the four prio arms at λ 0.3 and 1.0 with three CMA-ES starts per seed (cant started at 4, 32 and 64°), keeping the best training score. Prediction R1: with restarts the chosen cant at cant ≤ 80° is above 25° in at least 4 of 5 seeds at both λ, and cant wins the composite in 5 of 5 at both λ. R2: the no-cant arm improves by less than 0.5° from the restarts (its landscape is not the problem).

### Outcome of restarts.py (experiments/m2_polytope/restarts.txt, 60 CMA-ES runs, about 16 minutes; simulated)

  • R1 held for the angle, partly refuted for the win: with the best of three starts the chosen cant is 32–37° in 5 of 5 seeds at λ = 0.3 and 28–32° in 4 of 5 at λ = 1 (the fifth found only a 4° design from all three starts). Cant wins the composite in 5 of 5 at λ = 0.3 (mean −5.28, SD 1.67) and 4 of 5 at λ = 1 (mean −2.92, SD 2.56), not 5 of 5.
  • R2 refuted at λ = 1: the no-cant arm also has local optima. Its held-out RMS moves +0.17° at λ = 0.3 and −0.63° at λ = 1 with restarts, and its three training scores differ by up to 2.9° within a seed.
  • H1 (cant optimum 5–20°) is refuted under this objective. Pricing hover power shrinks the vehicle 5× in power, not the cant: 28–37° in 9 of 10 seed-arms from λ = 0 to 1 deg per 100 W, under both mixers. Milestone 1's reading ("the plateau is held up by the objective") was wrong in its attribution: what the free-power objective held up was the vehicle's size, not its cant.
  • Reading: on one shared thrust polytope a no-cant quadrotor tuned for attitude steps runs roll and pitch into saturation almost all the time, and an attitude-priority mixer then gives yaw nothing; cant moves yaw onto the tangential force, so the same mixer starves it less, and the cant advantage grows from 4.8° to 6.1°. Airmode recovers 1.6° for no cant and 0.3° for cant.
  • Limits: TWR fixed at 2 (where hover sits mid-range, the worst case for a hover-held mixer); λ is a modelling price, not a measured mission value; attitude steps only, five seeds, one κ, inherited mass, cost and motor fits; the CMA-ES landscape under the polytope is multimodal, so single-start results under prio are unreliable. Preliminary.
  • Next: milestone 3, motor catalogue data replacing the inherited fits, with TWR free under the polytope mixer and multi-start search as the default; then INDI and geometric baselines with 10 seeds.

2026-11-12 · Milestone 1: are the earlier cant results artefacts of the actuator model? (predictions written and committed before experiments/m1_allocation/run.py was run)

  • Why: the seven codesign-aero-cant* projects gave the attitude simulator two hand-written per-axis torque limits, roll 4 (Tmax − Th) r cos α and yaw 2 (Tmax − Th)(κ cos α + r sin α), with r = L/√2 and L the arm length. Writing out the full 6×4 map from rotor thrusts to body wrench (src/codesign_cant/allocation.py) and maximising each pure torque with a linear programme (thrusts in [0, Tmax], other torques zero, no net side force) gives, with Δ = min(Th, Tmax − Th) at hover or Tmax/2 with free collective: - roll 4Δ (L cos α − κ sin α)/√2: the earlier formula is right at TWR 2 apart from a small κ sin α term, but grows without bound above TWR 2 because it lets the low rotors go below zero thrust; - yaw 4Δ (L sin α + κ cos α): the earlier formula has a factor 2 where 4 belongs (it raises one diagonal pair without lowering the other) and the tilt moment arm r = L/√2 where L belongs (the tangential force acts at the rotor, a distance L from the centre). At TWR 2 it understates yaw authority by 2× with no cant and up to 2√2× with cant; - side force per unit roll torque sin α / (L cos α − κ sin α); codesign-aero-cant-poshold used tan α / r, √2 too large. - Closed forms agree with the LP to 1e-15 on 300 random designs (tests/test_allocation.py).
  • Plan: experiments/m1_allocation/run.py reruns the twrceil co-design (1920 evaluations, 5 seeds, 64 held-out episodes, c = UIUC median 0.081) with cant bound 0/35/80°, TWR fixed at 2 (as -propdata) or free in 1.5–6 (as -twrceil), and actuator model prior (the earlier formulas), hover (allocation, collective at hover) or free (allocation, collective free; TWR-free arms only).
  • Predictions: - P1 (reproduction): the prior arms reproduce the earlier results.txt numbers for the matching cells (twrceil: c 0.081, ceiling 6: no cant 13.65°, cant ≤ 35° 11.22°) to within 0.01°. - P2: with hover limits the no-cant design improves by more than 2° RMS against prior at TWR 2 (its yaw authority doubles). - P3: cant's advantage at least halves under hover against prior (paired mean, TWR 2, bound 80°: −4.6° in -propdata). - P4: cant still helps under hover (paired mean below zero, better in at least 4 of 5 seeds at bound 35°, TWR 2). - P5: the chosen cant falls: median under 30° at bound 80° with hover (36–41° in -propdata), because each degree now buys more yaw. - P6: with TWR free and hover limits the optimiser picks TWR below 2.6 on average (authority stops growing at TWR 2); with free limits it picks more than with hover.
  • Analysis: per-arm mean and per-seed held-out RMS; paired cant-minus-no-cant differences by seed with SD and sign count. Five seeds, no significance test claimed beyond the sign count.

### Outcome (experiments/m1_allocation/results.txt, 75 CMA-ES runs, about 9 minutes on 4 cores; simulated)

  • P1 held: the prior arms reproduce the earlier numbers exactly (twrceil ceiling 6: 13.65° and 11.22°; propdata TWR 2: no cant 17.06°, bound 80° 12.46°).
  • P2 refuted: with hover limits the no-cant design improves by 1.62° (17.06° to 15.44°), not more than 2°. Doubling yaw authority cuts no-cant yaw RMS only from 28.4° to 25.4°; yaw still dominates.
  • P3 refuted: cant's advantage does not shrink. Paired by seed, cant ≤ 35° minus no cant is −4.81° (SD 0.12, 5 of 5) under hover against −4.62° under prior at TWR 2. With TWR free it is −4.81° (hover) and −2.77° (free) against −2.43° (prior).
  • P4 held: cant helps in 5 of 5 seeds in every actuator model and TWR arm.
  • P5 refuted: the chosen cant does not fall. At bound 80° under hover it is 35–37° (median 35°), the same plateau as before; at bound 35° it sits on or near the cap (31–35°).
  • P6 held: with TWR free, hover limits make the optimiser pick TWR 2.0–2.3 (the earlier "interior optimum near TWR 5" disappears: above 2 the low rotors cannot go lower); free limits pick 3.3–4.7.
  • Found, not predicted: the cant optimum is held up by the objective, not the physics. The cant ≤ 35° design under hover, TWR 2, weighs 3.17 kg and costs $273 against 1.74 kg and $143 without cant (5-seed means), and 35° of cant needs cos(35°)^−1.5 = 1.35× the ideal hover power. The score is attitude error alone, with cost only as a $300 cap, so the optimiser spends mass and power on cant up to the cap. Any claim about the best cant angle needs an objective that prices power or endurance.
  • Reading: the earlier formulas were wrong in three places (yaw authority 2× to 2.8× too small, roll authority too large above TWR 2, side force √2 too large), but cant's attitude-error advantage, and its 35° plateau under this objective, survive the correction. What does not survive is the motor-headroom story of -twr and -twrceil: whether extra TWR helps depends on whether the mixer may give up altitude (free) or not (hover), which those projects did not model.
  • Limits: per-axis box limits (combined commands share one polytope, milestone 2); attitude steps only; inherited assumed mass, cost and motor fits; five seeds; one κ (the median). Preliminary.
  • Next: milestone 2, the attainable-torque polytope for combined commands and an attitude-priority mixer, plus an objective with an energy term (hover power or endurance at fixed battery), so the cant optimum is priced.

Paper, code and results

More in Agentic co-design of low-cost robots

All 28 Agentic co-design of low-cost robots projects →

Other research topics

Proper scoring & elicitation28Effort & peer prediction9Verification games & slashing19Detection & tolerances15Markets as routers13Prediction-market design9Paying contributors8Decentralised training systems23Digital twins & sim-to-real164LLM chip design loops40Robot manufacturing loops53Ambient AI assistant46Self-replicating manufacturing34Simulation-first design64Engineered clothing45Task-specific AI accelerators40Algorithmic chip-design loops42Open-source purpose-built robots50The software-engineering loop43Adversarial human imitation41Open-source residential construction41Autonomous multi-angle capture35Sensing, imaging and 3D33Energy supply and dense energy storage41Robot societies among people32Agentic co-design of low-cost robots28Laundry-hanging robots26Machine learning on qualitative data23Robots on the building site28Civilizations & group psychology3History & human sciences4

Produced with AI assistance (Claude). Results are from the code in this repository and are simulated or stylised unless the abstract says otherwise; nothing here implies affiliation with or endorsement by any lab, researcher or company cited. MIT licensed.