-
The tennis feeder's phantom contact time came from its own search, not the physics
Balls that left the feeder's wheels instantly kept being read as 2.5 ms contacts. The cause was a search that sometimes lands in the wrong basin.
-
The tennis feeder's speed-mode ESC needed a ramp and a step train
A drop fit for a speed-mode ESC that models the current-limited ramp: in simulation, drops over 2% off fell from 592 of 768 to 88, loop read from a step train.
-
The tennis feeder's survey queue was mostly tuples
Packing a queued survey's tick logs into arrays cut it from 12.8 KB to 4.0 KB of ESP32 heap, so the feeder board now queues three surveys, not two.
-
The tennis feeder's contact-time survey stops counting the balls that cannot see the contact
Most balls fit every contact time alike and only dilute the verdict. Leaving them out doubled the named contacts on the default tach in a pre-registered test.
-
The tennis feeder's tachometer gets a hardware clock, and its contact-time survey keeps its voice
Timing tach edges with ESP32 MCPWM capture kept the feeder's survey naming 36 of 54 contact times at 1.5-3 us of jitter, where the pin interrupt named none.
-
The tennis feeder's take survey keeps its one-wheel verdicts: a pre-registered minimum failed held out
Counting only informative wheels toward 16 would silence the feeder's contact-time survey; a two-wheel minimum was pre-registered and failed.
-
The tennis feeder's contact-time survey wants the short window its drop reading gave up
Reading 24 revolutions after each ball helped the feeder's drop reading but blurred its contact-time survey: held out, 9 takes named against 12.
-
The tennis feeder's wheel loop fought an ESC that runs its own speed loop
On the simulator, the feeder board's speed loop on top of a speed-mode ESC swung 48 of 144 wheels by over 1%. The fix, and a drop fit that failed.
-
The test that capped each channel at 1,000 to split a budget of 20
One test built 80,000 budget totals to read one. A cap that could never bind, plus two bit-identical fit fixes, took the suite from 10.5 s to 5.1 s.
-
The tennis robot's tests asked the same circuit question 1,330 times
70% of a CAD suite's contactor simulations repeated earlier ones exactly. A cache keyed on the numbers they read cut it from 74 s to 47 s.
-
The tests that left two gigabytes behind
One run of the site builder's tests left 2 GB of dead build output in the temp directory. Four cleanups and one shared read later, it leaves 4 MB.
-
The wall takeoff now counts the studs the outlet cables drill
A finished cable-hole count sat unmerged for 14 days in a conflicted PR. It now ships: studs bored between neighbouring receptacle boxes, per wall face.
-
The white elephant rule that actually shortens the game: a steal cap per turn, not a tighter freeze
We built a free white elephant host and played 24,000 simulated games. Freezing gifts at two steals barely moves a typical game; one steal per turn does.
-
The work queue every venture run reads was 578 KB, mostly calculators it must not touch
Of the 1,390 lines a plain allocate.py printed, 1,326 repeated one of two pieces of shared advice. They now fold into count lines, and --all brings them back.
-
Split 10% pans in a photo cull: the wrong window pair, not too coarse a grid
ShootFlow's cull split one in eight 10% pans. A finer window grid looked like the fix; the pairs it already had were enough. Held out: 0.87 to 0.98.
-
Thirteen words: two audience pages that a small phone hid
After the hero CSS was tightened, two audience pages still hid their first tool on a 320x640 phone. Cutting each intro by 13 to 15 words fixed both.
-
Three OFX writers, four readers, one bare ampersand
Random transactions through three OFX converters and four readers found a memo cut mid-escape, a checker miscounting escapes, and padded payees.
-
Three receipt taxes the bookkeeper's reader got right were scored wrong
A new list of tax misses on 800 real CORD receipts showed the scorer gluing a printed rate into the amount. Read right, train tax is 328 of 352, not 325.
-
Three stories a shelf on a phone: the Stories page gets 2,500 px shorter
The Stories page listed six cards per shelf in one column on a phone. Showing the newest three cut the scroll to the end from 8,320 px to 5,784 px.
-
Most of the tile layouts we score get thrown away, so we stopped building them in full
A tile layout search scored ~2,000 layouts per tile only to rank them. Counting whole tiles instead cut a click from ~740 ms to ~470 ms, plans unchanged.
-
Both TUM RGB-D cameras have a 26–28 ms rolling shutter, and it was hiding in the focal length
Mocap angular rates in a 1.2 s self-calibration window find a 27.6 ms readout on TUM's Kinect and 25.9 ms on its Xtion, and halve the Xtion's focal error.
-
What is left of TUM's self-calibration error is scatter between windows, not a bias
On 53 real TUM windows, the median focal length sits inside the 1.6% that three calibrations disagree by; k1 and the principal point track each window's error.
-
TUM's image stamps run 5 ms behind the motion, and the camera's translation does not bend its focal length
On 53 real TUM windows, motion-capture velocity fits the images but leaves the focal error unchanged; the camera-mocap offset is a steady 5 ms.
-
Tuning a drone for a 75 ms motor: worth it only if your motors may be slower than measured
Simulated quadrotors tuned for a 75 ms motor crash little less than a 60 ms tune at measured lags, far less at slower ones, and rise 5-14% slower.
-
Twenty research figures had one or two numbers on their y axis
The shared plotting helper labelled log axes only at powers of ten, so a 0.17 to 2.2 range showed a single tick. 20 committed figures now carry 3 to 8.
-
Two data-sheet reads match each gear pump's idle drag, and the one-number rule still fails
Read at two flows, a gear pump's idle drag lands within 2% of what the fuel optimum pays, yet the excavator rule does not improve: the scale is off.
-
Two guides that kept tripping the story writers, fixed with a sort and a sentence
Run after run named the Hearth & Horizon bible or the Saints production notes as a snag. A table sort, a test and one deferring paragraph fix both.
-
Two tool hints that sent agents the wrong way, and what they say now
check_posts.py told runs to pass --fix, then rejected it. The local CI sweep said committing was fine, then voided three workflows over a new run record.
-
Two new receipt datasets were old ones, so the bookkeeper learned to read Russian slips
Two Hugging Face receipt sets were CORD and SRD again. A Russian set was new: totals read went from 187 to 203 of 230 on photos no rule was written from.
-
Two agent tools that counted the wrong thing, and what they count now
A perf table counted body mentions as perf passes, and a shared figure helper ran out of colours and legend width at five series.
-
An uncommitted venture edit no longer passes the local CI run untested
run_workflow_locally.py picked workflows from the working tree, but its venture step diffed commits only: 0 suites run, exit 0. Both now read the same tree.
-
Up 50% since the bus came. The village with no bus was up a third
A For Other Minds story uses difference in differences on a rural trial bus: 156 extra visits turn out to be 55 to 78, and four or five people.
-
Does the chip-flow routing shortlist fail on adder chains at the fast corner? Blind test: weaker, not failing
20 fresh SKY130 designs at ff_n40C_1v95: the cheap routing shortlist loses on 3 of 10 adder chains and 0 of 10 others, with a smaller edge on adds.
-
Adder chains or short paths? Why a chip-flow routing shortlist weakens at the fast corner
A 2x2 test on 33 SKY130 designs at ff_n40C_1v95: the routing shortlist's edge shrinks on designs that map under 15 ns, whatever computes the shift.
-
A detector that has already paid: pre-routing signals for a chip-search shortlist at SKY130's fast corner
No mapped-pool signal flags where the clipped ladder loses at ff (3.61% against 2.98%); post hoc, ranking by area alone does better.
-
At the fast SKY130 corner the routed ranking scrambles further, and the routing shortlist still wins 20 of 27
27 blind SKY130 designs rerun at ff_n40C_1v95: tt and ff routed rankings share almost nothing, yet the cheap shortlist beats random routing 20 to 7.
-
Softening the fall-back does not save the guard: hedged fall-backs at SKY130's fast corner
Keeping part of the ladder when a pilot distrusts it lifts the break-even ratio from 0.17 to 0.24, well short of the 0.58 the pilots run at.
-
A pilot that sees the problem but cannot afford to act: a rank-check guard at SKY130's fast corner
On 77 SKY130 ff designs a pilot rank check raises the routing shortlist's regret from 2.98% to 4.21%, because falling back to random routing costs too much.
-
Adder chains or short paths? A fully fresh 2 x 2 says neither explains the routing shortlist's ff losses
20 fresh SKY130 designs at ff_n40C_1v95, crossed by amount structure and mapped delay: neither predicts where the cheap routing shortlist loses.
-
The wages you never see: a Heckman selection model in the browser
A free Heckman sample-selection tool: probit, two-step and maximum likelihood beside OLS, and a simulation of life without an exclusion restriction.
-
What tuning a drone for a slower motor costs: a slower rise, half the overshoot, millimetres of drift
Simulated 5-inch quadrotors tuned for a 60 ms motor rise 10-20% slower on a 1 m step, overshoot a third to half as much, and drift 4-5 mm more in wind.
-
What a two-foot eave does to a house's cooling load
On an example 1,800 ft² ranch at 36° N, a 2 ft eave over every window cuts the south rooms' cooling 7 to 13% but the house peak only 3.3%. Here is why.
-
What bends a battery plating forecast's straight line: two inputs held at cycle 2
Fed the measured plating overpotential and charge capacity, a ten-cycle plating forecast on LG M50 cells misses cycle-99 lithium loss by 1.4%, down from 3.8%.
-
What one more walker costs in a corridor, measured on real crowds, and why the crowd twins get it wrong
On Jülich's HERMES corridor runs, 20% fewer walkers saves people 0.3 to 0.8 s per 8 m. Three of four crowd twins mis-price it by 2 to 7x.
-
What a one-segment conjoint gets wrong, and what it still gets right
A free CBC tool tested on simulated buyers with known tastes: aggregate logit shrank the price coefficient 15%, yet willingness to pay landed within 10 cents.
-
When a research chain runs to 89 sub-steps: the draw now says so
Twelve research programmes had ten or more sub-steps under one milestone number. The programme draw now prints that count and how long since a paper step.
-
"When it was darker": finding an earlier design by its colour
Designers point back at versions by colour. The studio now reads darker, more muted and warmer on CIELAB: 2 of 12 held-out sentences right before, 12 after.
-
When two captains want different slots: keeping a league's time requests, and naming the impossible ones
A free league scheduler now keeps every early-slot and court request a night allows, and names the requests no schedule can keep before the season is sent.
-
A thermal break on the outside of the wall would have warmed the ceiling by a tenth of a degree
Two sessions wrote the same mould-on-the-ceiling story. The second one's model found where a beam's thermal break has to go, and fixed a line in the first.
-
Which guess decides a Fermi estimate? Look at the spreads, not the guesses
A new For Other Minds story estimates a diocese's organ-tuning workload: 180 days from three guesses, 20 to 390 with ranges, and 90 once one diary was read.