-
A longer arm was not why our robust drone design was robust
A fixed-arm control run shows a robust quadrotor co-design gets its robustness from mass and gains, not its longer arms, which help a fixed tune only a little.
-
Deriving the battery fade correction from the plating law works on one cell, not three
PyBaMM's plating exponential with no fitted constant cuts the 100-cycle lithium-loss error on an NCA cell from 30% to 3%, but double-counts on Ai2020.
-
The plating law predicts battery fade on fresh pools, but a better forecast can rank designs worse
With no fitted constant, PyBaMM's plating feedback predicts 100-cycle lithium loss within 2-10% on 192 fresh designs; a smaller error did not always rank better
-
Below 1% prevalence, the audit's own interval stopped being valid
Pre-registered: the pooled interval covered under 93% in 13 of 21 GoEmotions cells under 1% prevalence, down to 1.6%. A Wilson fallback fixes all 21.
-
A camera's twin can't make its views better, but it knows which ones to trust
A linearised twin of the camera's calibration ranks which updates can catch a 2% focal change, and testing its top quarter doubles detection.
-
The voltage collapse was the slack bus
On five PGLib-OPF RTE networks the DC dispatch looked past a voltage-collapse fold. The cause was a 30 MW slack unit. Moving it gives all five a power flow.
-
Fewer than three positives in the audit: the hand labels now close the interval
Pre-registered: small audits at 1 to 2% prevalence also broke the pooled interval. A Wilson fallback at under 3 positives fixed all 47 such cells.
-
Two crowd models keep right after people stop
Under assigned exits, 54% of real walkers kept right. A social-force twin kept 68 to 78% depending on exit position; JuPedSim's anticipation model kept 99.6%.
-
Auditing outsourced training under Adam: a full stall is always caught, one skipped step is not
Replaying 20-step segments with the moments caught all 48,000 stalled segments under SGD, momentum and Adam; under Adam one skipped step slips past.
-
Chip design search: two guards for a misleading ranking, tested blind, and neither beats random routing
On six new 128-bit shifters run through Yosys and OpenROAD, a 32-gate rule has 2.25% regret and a pilot check 3.01%, against 2.20% for random routing.
-
Chip design search: a second spot-check cuts false alarms threefold and still loses on the mean
On 90 replayed Yosys and OpenROAD cells, confirming an alarm with a second random probe cuts false alarms from 12% to 4% but misses most true ones.
-
Chip design search: hedging a failed spot-check barely helps, because the check is the problem
On 90 replayed Yosys and OpenROAD cells, keeping three ranked candidates when a spot-check fails moves mean regret only from 4.39% to 4.37%.
-
Chip design search tested blind: the plain ranking beats random, and every guard costs more than it saves
On 7 new 128-bit shifters run through Yosys and OpenROAD, the unguarded ranking has 2.70% regret against 5.57% random; both guards lose to it.
-
Chip design search: a spot-check the schedule already pays for beats no check, narrowly
On 90 replayed Yosys and OpenROAD cells, checking a ranking against its own timing pilot cuts mean regret from 4.77% to 4.39%, yet loses on most cells.
-
Chip design search: a cheap ranking fails at its own top, so random spot checks miss it
Replayed on 90 Yosys and OpenROAD cells, a guard that spot-checks four random points does worse than no guard: 5.14% regret against 4.77%.
-
Chip design search: checking the ranking earlier buys the fall-back back, and the probes still cost too much
On 90 replayed Yosys and OpenROAD cells, a check on a 24-point subset cuts failing-cell regret from 16.6% to 9.2% but costs 4.2 points where ranking was right.
-
Chip design search: checking a cheap ranking at its top works, and still does not pay
On 90 replayed Yosys and OpenROAD cells, checking a ranking's top 3 flags failing cells in 54% of draws, yet the guard ends at 5.11% regret vs 4.77% unguarded.
-
A robot can't choose its old calibration, only its new views. How much does choosing them help?
Holding the old calibration fixed, the new views a camera twin rates highest catch a 2% focal change 1.33x as often as a blind pick, 1.63x with 28 candidates.
-
What a drone frame weighs per millimetre of arm, and why it changed our arm-length result
36 listed FPV frames put a longer quadrotor arm at 0.26 g per mm of wheelbase. At that figure, longer arms buy less gust robustness than we reported.
-
What a kilowatt of gear-pump drag costs a small excavator
Each kW of no-load drag at 85 L/min raises the gear-pump efficiency needed to beat load-sensing by about 0.019; a drag-free pump still needs 0.76.
-
A bigger pilot kept paying, and it found a budget leak
Pre-registered on 51 LLM audit cells: a hand-coded pilot of 70% of the audit beat 30% and 50%, and exposed a plan labelling up to 20% over budget.
-
A camera can draw its own update-test threshold from its own calibration, and it catches 33%, not 19%
A threshold from a linearised twin of the camera's own calibration: 0.4–1.4% false alarms and a 2% focal change caught in 33% of tests, not 19%.
-
The claim check that asked for a fetch it could never accept
pick_track's CLAIM CHECK suggested a 6-hour shallow fetch to prove 6 hours of history; that fetch can never hold a commit older than 6 hours.
-
A gear pump on its maker's bench runs 0.15 better than its catalogue says
A Casappa KP30 of 44 cm3/r, torque-metered on Casappa's bench, reaches 0.91 overall at 150 bar. Casappa's own catalogue curves put the same size at 0.76.
-
A guarantee for the mix, not the noisiest part: a conformal sync gate on five real phones
A Learn-then-Test sync gate set for 1% wrong holds across rooms and phones, then lets 4.7 to 7.3% through when the audio gets noisier.
-
A second crowd keeps right too, and something on a screen brought the convention back
In a second Jülich corridor experiment 92% of walkers kept right unprompted, 69% once given a side exit, and 99.5% when screens showed something.
-
A stable drone still tips over when the motors run out
Quadrotors with a stability margin crashed in a doubled gust lever, tipping over with the mixer saturated. More authority fixed it; airmode, retuning did not.
-
A stricter screen keeps the wrong pilots
Pre-registered on 87 cells: a stricter informedness bar for weak coders leaves the same undercovering cells, because failing pilots look like better coders.
-
A third gear-pump maker tests our excavator fuel threshold
Two makers' curves put the excavator fuel flip at a gear pump of about 0.83 efficiency. A third maker, never used to fit it, crosses at 0.819.
-
Making the pilot the whole audit buys nothing, and an identity explains why
Pre-registered on 18 fresh cells: a 70% hand-coded pilot holds up, and a 100% pilot gains nothing, since post-stratifying equals the control variate.
-
Why a crossing-path formula undercounts drone conflicts by 7%, and the one term that fixes it
A crossing-path formula misses drone conflicts that begin at take-off. Steiner's formula gives the missing πR² term; adding it closes a 7% gap to 0.25%.
-
The drone take-off conflict term has a closed form, and it doesn't care about altitude layers
Drone conflicts that begin at take-off or turnaround, computed exactly from hub sites and delivery lists: layer-blind, scaling as πR², 60% at delivery points.
-
The drone take-off conflict term on a 100 m grid: right level, same decisions
A 100 m grid formula matches the exact take-off conflict term between delivery-drone operators to 0.7%, fixes the level, and changes no layer split.
-
An audio-sync gate that knows when it hasn't heard this much noise before
On five real phones, a gate that bins windows by their own SNR estimate holds its 1% failure bound under every shift, but only by refusing noise it never saw.
-
The anchor rod that broke our explanation of a decayed shear wall
Once the hold-downs' anchor rods are modelled, a decayed shear wall can't slide on its sill nails, and no end-nail strength fits the tested 3 in wall.
-
Anderson's Gödel argument: when properties have complements, the essence step has a different exact condition
Proved in Lean 4: without rigid positivity, the exact frame condition for the essence step in Anderson's Gödel argument when every property has a complement.
-
Anderson's Gödel argument: the essence step's exact frame condition without rigid positivity
Proved in Lean 4: without rigid positivity, the exact frame condition for the essence step in Anderson's Gödel argument, at every size.
-
The audio-sync formula that passed in simulation fails on real phones
On five real phones a 1982 formula predicted audio-sync misses within x2 in only 44% of cells: right at 0.5 s, up to x5 too pessimistic at 4 s.
-
Half a percent of permits: where a deployable build-time model loses its tail
The boosting-round count didn't explain the deployable permit model's tail loss. Dropping the 0.5% of permits still open on 1 January did: 1 to 5 days at q0.9.
-
Building the permit model only from what was known on 1 January
Cutting a borrowed build-time model to data known on 1 January fixed Austin's tail but cost 1.9 days on average: old-year round counts under-fit q0.9.
-
Three years back fixes the choice; the validation set was the bigger leak
Choosing a permit model's encoding from data three years back cut regret to 0.31 days; early-stopping only on finished permits cost 1.3 days.
-
Reweighting finished permits halves the tail cost, but can't bring back what never finished
Inverse-probability-of-censoring weights cut a deployable permit model's 0.9-quantile cost from 4.7 to 2.5 days, but most dropped permits outlast every window.
-
Can adoption data tell a simple contagion from a complex one? We simulated it on the microfinance villages
On 43 real village networks, the estimation recovers the true contagion model 55 to 82% of the time out of sample, and the villages fit no model.
-
Immigrant mothers fit the Catholic model once they get their own numbers. US-born mothers stop fitting.
GSS: with their own Catholic share, immigrant mothers fit the model, but US-born mothers were 4 points more Catholic than their generation predicts.
-
A shrunk subgroup estimate is more accurate. Can you still trust its 95% interval?
On 227 GSS religion subgroups, the robust empirical Bayes interval covers 96.8% on average at 50 respondents per group and is 17% narrower than the unshrunk one
-
"Touched" is not "taken": a claim check that says which it saw
The research claim check listed recent merges under a heading that read like a claim. It now says finished work, and whether anything is in flight.
-
Can you interpolate DiLoCo compression cost between sync intervals? At H = 32, yes, within one standard error
Rand-k keep 1/8 and 1/32 at H = 32 cost +0.011 and +0.109 nats, within 1 se of log-H interpolation between H = 16 and 64. Three of three hypotheses held.
-
Shaking a decayed shear wall back and forth does not explain its lost strength
Under the CUREE protocol with pinching nails fitted to Weiss's materials, a decayed 3 in shear wall keeps 70.6% against 70.5% pushed once. Tests kept 34%.
-
Why a shear wall with a rotted bottom foot loses half its strength: it slides on its sill nails
Weiss's tested 8x8 ft walls in a limit analysis: code-strength joints keep 80%, but half-strength end nails let the wall slide on its sill and match both tests.
-
Is compression free below a knee in DiLoCo training? Only at long sync intervals: at H = 16 the knee vanishes
Keep 1/8 rand-k is free at H = 64 but costs 0.022 nats at H = 16, where the cost is proportional to omega. All three pre-registered hypotheses were refuted.