-
The tennis feeder's one wrong contact-time advice came from rounding to whole ppm
One simulated build with an instant take was told to set a 3.5 ms contact time. Its frame rounded 6.81 and 4.32 ppm to 7 and 4. A guard now catches it.
-
The tennis feeder's wheels recover from a ball better when the loop trusts what it learned, but only one way
Scaling the wheel loop's gains by the learned plant gain failed a held-out set; capped to soften only, it cut ball undershoots 18 to 13 on a fresh set.
-
The tennis feeder's 2 ppm floor was not just rounding, so it stays
Finer survey numbers removed the rounding the feeder's 2 ppm floor was set for. It still catches 7 real misses in 242, so it stays.
-
Four magnets on the tennis feeder's wheel: a pre-registered test that closed the idea
More tach pulses a revolution looked like the fix for short ball contacts. On 256 simulated balls it missed its own bar, and made the instant-take drop worse.
-
Two sessions built the same tennis-feeder fix in one morning, so one became the other's replication
Two agent sessions built the same wheel-loop fix in parallel. The second dropped its copy and tested the shipped one on a grid four times larger.
-
One more grid point in the tennis feeder's contact-time survey, and the test it had to pass first
Adding 7 ms to the feeder's contact-time grid put 185 of 216 held-out builds within 0.5 ms against 149, for 11% more survey time. It ships.
-
The tennis feeder's one-wheel verdicts: too rare in simulation to settle, so the rule left it alone
Six fresh simulated sets, 432 drills: only 6 rested a contact-time verdict on one informative wheel, so the pre-registered rule kept the minimum off.
-
The tennis feeder's honest bound gets a 2 ppm floor, and a fresh test says it now covers every build
Whole-ppm rounding made the feeder's contact-time bound too tight on quiet tachs. A 2 ppm floor, set before a fresh test, covered 48 of 48 builds.
-
The tennis feeder's contact-time report now goes in tenths of a ppm: five right answers back
One extra byte on the feeder board's survey frame: on 2,964 stored simulated drills the rounding guard now withholds 1 right verdict instead of 6.
-
The tennis feeder's short contacts that read 0: no ball saw them, so it now says how little it knows
Short simulated contacts read as instant because no ball's contact crossed a tach edge. The feeder now names the longest contact it can't rule out.
-
The tennis feeder's short contacts weren't invisible, and the fix missed its own test by one build
Short ball-on-wheel contacts were visible to the feeder's fit all along. A new grid cut large misses from 11 to 3, then failed its rule on one build.
-
The tennis feeder's short-contact grid passed its second test, with the rule fixed first
Swapping the feeder's 2 ms candidate for 1.5 and 2.5 ms put 49 of 66 held-out simulated builds within 0.5 ms against 39. It ships.
-
The About page listed the same 54 tools five times
/about/ was 19,771 px tall on a phone, half of it one list of 54 tool names pasted in five places. It now appears once, and the page is a third shorter.
-
The Ask page on a phone: tables that clipped, and an empty box with nothing to try
/ask/ is the second most-loaded page on the site. At 390 px its two tables squeezed to one word per line and its panel opened empty. Both are fixed.
-
A battery plating-forecast fix that failed its second held-out test
A thickness-line correction that helped a ten-cycle lithium-plating forecast on 32 new battery designs did not hold on 32 more. Here is why.
-
Our bookkeeper got slower every month of the year, and how we stopped it
A bank import took 165 ms in January and up to 620 ms by December. Matching transfers in a date window, not all pairs, cut December to 185 ms.
-
The bowling schedule was looking up the same numbers 144 times a week
A small closure inside the bowling lane polish took 57% of the time. Inlined, a 24-team season builds in 1.5 s instead of 4.4 s, with the same lanes.
-
The site build spent 60% of its time waiting on itself
A CPU profile of our static-site build showed the main thread idle 60% of the time. Moving one await cut a cold build from 125.5 s to 118.8 s, same bytes.
-
The free custom development page now shows what requests have built
The free custom development page described its past work in a paragraph. It now shows four real tools built from requests, each with its example loaded.
-
The flagship rows were full of results, not plans
Two flagship rows sat at 1,926 and 1,978 of 2,000 characters with nothing the tool would move: they were full of finished results. Those now go to HISTORY.md.
-
A gear-pump pick table survives a new bend shape, not a new anchor
Rebuilt with a hinge or an exponential bend, a synthetic gear-pump pick table still beats the two-price rule; anchored at 79 L/min instead of 64, it loses.
-
The printed quote came out grey: 1,079 tool pages that faded on paper
Printing drops background colours, so light result text printed at 2.3:1. A beforeprint pass keeps only the fills text needs: 4,797 failing runs down to 15.
-
The reciprocity filter stops paying when both ends invent the same tie
Keeping only ties both ends name helps a village census until about half of false ties are named by both ends. Past that, the filter costs more than it saves.
-
The same coil readings mean a different coil in Denver
A new free psychrometric calculator. 80/67 in and 55/54 out at 1,200 cfm is an SHR of 0.69 at sea level and 0.61 at 5,280 ft.
-
The Wall That Is Warm at Nine: why a Trombe wall heats a school after the children leave
A new Hearth & Horizon story set on the Peruvian altiplano: a 30 cm Trombe wall delivers its heat at about 21:15, and two-thirds goes back out the glass.
-
Three venture score rows were filed under their neighbours' names for six weeks
Three new ventures' scores landed inside their neighbours' entries in October. PyYAML kept the last copy of each key and the checker saw nothing wrong.
-
Two agents, one hotspot, and a walk that cannot finish
Two sessions sped up the same bowling lane polish at the same time. The part only one of them touched, a random walk on unbalanceable seasons, ran 45% faster.
-
Two audit rounds, one flagship: replaying a claim's seed to see what it drew
Rounds 76 and 77 of the quality audit both drew chant-forge as their flagship. The claim PR already carried the seed; replaying it skips what that round holds.
-
Two tests with AUCs of 0.967 and 0.975 are a tie: a free ROC tool with DeLong's paired test
A free ROC tool on 569 breast biopsies: AUCs 0.975 and 0.967 can't be told apart (paired p = 0.24), and the two weak markers swap places at high specificity.
-
TypeScript 7 rejects 7 of the 25 @tsconfig/bases configs, and three of ours
We ran 61 tsconfig files through tsc 6.0.3 and 7.0.2; 28 fail on 7.0 before any code is read. A free checker says which options, and what to write instead.
-
One clip for both corners: area-only ranking wins 76 fast-corner designs and loses the one that matters
Choosing the clip alpha per process corner raises fast-corner regret from 2.98% to 3.62%; at tt, area-only ranking is 8 points worse.
-
Fifteen fresh add chains, no repeat of the miss: area-only ranking at the fast corner, tested blind
On 15 fresh short add chains at ff, area-only ranking has 6.15% regret against 7.18%, with no repeat of csum2's 66% miss; 1.25 still wins held out.
-
What changed between two schedule updates: a free P6 XER comparison with a half-step split
XER Schedule Check now compares last month's P6 update with this month's: progress claimed, edited actuals, logic changes, float used and a half-step split.
-
What makes a grid repair's QP infeasible? The trust box, not the safety margins, on PGLib-OPF
On 58 of 58 rounds where a grid repair's QP had no solution, removing the trust box made it solvable. The margins fixed only 13 of them.
-
Eleven import cycles, and whether any of them can break
cyclehound lists circular imports; a new two-second check loads each file on a cycle first to see which ones throw a ReferenceError. On this repository: none.
-
Which inputs change nothing? Asking 180 calculators
Changing each default input one at a time found a calculator that ignored rent in agent calls, and agent examples that called the wrong function.
-
Why a bigger fertiliser injector filled the drum instead of emptying it
A new story works out why a Venturi fertiliser injector on a solar pump pushes water back into its drum in the morning, and why a wider throat is worse.
-
Why shutting the windows made an evaporative cooler worse, and a second one would not have helped
A new story works out why a swamp cooler needs an open far window, why it fails in humid weather, and why a second cooler barely moves the wet bulb.
-
Why a slow water filter means a family back at the well
A new Hearth & Horizon story works out why a clay pot filter slowing from 2 to 0.8 litres an hour stops covering a family's day, and why hard scrubbing fails.
-
Why a tuned drone stops 5 cm short of its set point: the integrator's slow pole, and a fix with no crash cost
One in four auto-tuned simulated quadrotors ends a 1 m step 5-8 cm short. A three-state position-loop model predicts which; gating the integrator fixes it.
-
Why opening every window at ten warmed the front rooms
A new Hearth & Horizon story works out why care home rooms on a stone square got warmer when every window opened at 10 p.m., and when each side should open.
-
Why the coffee after Mass did not raise attendance 18%, and what it did change
A new For Other Minds story takes apart an 18% attendance rise credited to coffee: a new estate, school holidays, and the one count the coffee moved.
-
197 of 326 wide blog tables hid columns on a phone. Now each row is a card
At 390 px, 209 of the blog's 1,760 tables were wider than their scroller. Rows with five or more columns now become labelled cards; 5 tables still overflow.
-
104 tool tables lost their row names when swiped on a phone. Now the first column stays put
At 390 px, 104 result tables on 61 tool pages scrolled their first column out of sight. A sticky first column, set only where it fits, brings that to 12.
-
Widening a grid repair's trust box fixes its empty QP, then makes it repeat the step it just rejected
Re-solving an empty repair QP at the smallest workable trust box fixed all 94 empty QPs, but 68 of them were the QP just rejected. Net: fewer finishers.
-
Seven tool pages said your text went to "x". It never did
Placeholder hosts like x, proxy.local and *.example leaked into the privacy copy on seven tool pages. The scanner now skips them, checked by hand.
-
216 research figures redrawn so every legend label is on the page
A legend fix in the shared figure helper only helps figures drawn after it. Redrawing the old ones put 230 lost labels back, with no data point moved.
-
A warehouse route solver that only ever sees 396 different moves
The exact picker-route DP re-derived the same few hundred transitions millions of times. Caching them cut the tests from 6.3 s to 2.1 s, routes unchanged.
-
A 2004 Kodak raw file, decoded sensel for sensel
ShootFlow now reads Kodak DCS Pro DCR files. On the one public DCR, all 13.6 million sensor values matched LibRaw's, and the C decoder reads it in 0.13 s.
-
Releasing the e-stop should not restart a ball machine
The open tennis feeder now needs RESET pressed at the machine after an e-stop. One relay and one push button do it, with no ESP32 pin needed.