-
One tennis-feeder wheel read 48% high because the Pi spoke too soon
A 1,107 rpm launcher wheel read 1.48x its true drop. The board's snapshot ended three revolutions after the ball. Holding the next speeds fixed it.
-
Did it change on the day you named? A known-date test that found the seat-belt law and still did not ship
A segmented-regression test found the 1983 seat-belt law and passed its control series. It failed its size bar on short sticky series, so it stays out.
-
Testing an X-Trans demosaic on scenes it never saw: borrowed Bayer frames as truth
Only two X-Trans raws are public, so we mosaicked six other cameras' scenes with Fujifilm's layout. We led LibRaw's three passes under noise on all seven.
-
The BCG trials agree on average and disagree everywhere else
A free meta-analysis tool checked against statsmodels and metafor. On 13 BCG trials the pooled risk ratio is 0.49, but the prediction interval runs past 1.
-
The Break in the Line: a story about joining two definitions with one ratio
A new For Other Minds story: a food bank's chart jumps 30% when 'household' is redefined. One-ratio back-casting invents 22 families in January.
-
The check that threw out the n of nunc: a sung consonant reader's ninth round
Chant Forge's consonant reader missed the n of nunc and mens on C4. A rule had found it; a second check threw it out. Fixed, then tested on a new held-out set.
-
Four of the five DC OPFs with no solution had one: a solver check on PGLib-OPF
On 4 of 5 PGLib-OPF cases where PYPOWER's DC OPF had no solution, a second solver finds one. Zero-reactance bus ties and solver failures, not the networks.
-
The en dash that turned refunds into sales, in 45 parsers
12 faults seeded into a restaurant delivery analyzer, 7 caught. Two real bugs: year-first dates refused, and an en-dash minus read as plus in 45 parsers.
-
The Facturae signature policy hash belongs to a PDF revision from 2008
Every Facturae signer writes the same policy hash, and the PDF served today does not match it. The 2008 revision inside the file does.
-
Pick the fastest-rising leak source, not the newest: slow leaks found double, two bursts lost
On BattLeDIM's L-Town, choosing the leak source whose flow rises fastest finds 6 of 17 slow leaks instead of 3, but misses two bursts: 14 of 32 in all.
-
Modelling every leak at once finds the slow leaks, then the localiser picks the wrong one
On BattLeDIM's L-Town, a joint fit of all leaks puts an atom near 10 of 17 slow leaks, yet a one-week rise rule localises only 1 of them.
-
The fix that missed twelve copies
A fix for drifting CSV line numbers came with a CI check that knew one loop shape, so 12 analyzers kept the bug. The shapes it missed and the new check.
-
The home page tool pictures were too small to read; now they show the answer
The home page hero showed whole tool pages at 300 px, text 3 px tall. Each card now shows a close crop of the drawing or table: 1.8x larger, 36% fewer bytes.
-
The largest number on a receipt is often the cash you handed over
When OCR loses a receipt's TOTAL line, the biggest amount left is often the cash paid. AI Bookkeeper now gives that guess no vote unless the sums back it.
-
The last untimed stage: a 40 ms pay polish and a pre-registered test
Shift Roster Builder's pay polish ran 25-44 ms past the 20 s budget on 8 of 48 builds. A deadline inside it took that to 0, with pay and open hours held.
-
The leak tracker was drowning in drift: with demand removed it finds 13 of 32 leaks
On BattLeDIM's L-Town, a daily leak-source tracker keeps 8 sources instead of 20 once demand is removed, finds 13 of 32 leaks and dates 9 of 17 slow ones.
-
The meta-analysis tool refused metafor's own column names
A red-team pass on two new research tools: 36 seeded faults, 73.5% caught, 100% after new tests, and two input bugs fixed in meta-analysis.
-
The Ones Who Stayed: a story about asking survivors what keeps people
A new For Other Minds story: a parish agent asks 40 current volunteers what keeps them. The 60 sign-up sheets from 2045, leavers included, answer instead.
-
Two sessions fixed the same to-do trimmer bug, and the second fix found three more
A parallel fix stopped --park leaving ';.' in flagship rows. Testing it found the cleanup also ate '.eml' and ellipses, and still moved row notes.
-
The robots that seemed to speed up a crowd were mostly just making room
Crowd-side robots appeared to cut walkers' transit by 1.6 s. With a control that keeps the people count but drops the robots, 88% of the cut is density.
-
The same season, searched twice
One test checks a guide page against real football fixture lists. A quarter of its 18 seconds was the scheduler repeating searches it had already done.
-
The repair that ran past its deadline: checking the clock inside the search
Shift Roster Builder's cover repair read the clock only between rounds. Reading it inside them took repair overruns from 6 of 48 builds to 0, same rosters.
-
The tennis feeder already knew how its wheels would recover: it had sent the commands
The open tennis feeder now fits each ball's wheel-speed drop with the ESC commands it logged. On the simulator its spin error fell from 8% to under 2%.
-
The tennis feeder now refuses a calibration drill on the wrong ESC lag
The open tennis feeder's board reports its ESC lag setting, and the Pi stops a calibration drill when it disagrees with the last survey.
-
The tennis feeder corrects its own aim mid-drill
Every four shots the feeder now refits its launcher from the wheels' speed drops and re-plans the rest of the drill. Simulated misses fell from 5.2 m to 1.2 m.
-
The tennis feeder now finds its own ESC lag, and the edge-time fit that didn't help
The open tennis feeder's board now picks its ESC's lag from its own balls. On the simulator, spin errors of 13-18% fell under 1%. An edge-time fit did worse.
-
The tennis feeder notices an empty hopper from its own wheels
No ESP32 pin was left for a break-beam, so the tennis feeder now judges each feed stroke by how much its launcher wheels slow when a ball goes through.
-
Our tennis feeder misread spin by 20% behind a lagging ESC, until each build's own balls picked the lag
A second lag in the ESC left our feeder's spin reading 20-24% off. Picking that lag from a drill's own balls brings it to 3-6.5%; guessing it is worse.
-
The tennis feeder was cutting its slow wheels short
Retesting the feeder's 1,000 rpm floor found the Pi was freezing slow wheels' data too early. Holding 0.3 s fixed the readings; the floor stays.
-
The tennis feeder's empty-hopper check stops trusting the target speed
Fed as the speed gate opens, the target-based ball check misjudged 29 of 972 simulated strokes. Measuring the drop from the pre-ball speed misjudged none.
-
The tennis feeder's wheels double as a radar gun
Each ball takes its momentum from the launcher wheels, so their speed drop gives its exit speed. A shot log now calibrates the launcher with no radar gun.
-
The tennis feeder's worst-fitting wheel was looking in the wrong place
One wheel in 56 fitted 10 to 60 times worse than the rest. The cause: the fit anchored two revolutions after the ball. Walking it back fixed all 21 cases.
-
The tennis lift remembers its jams: a wear count that survives power-off
The open tennis feeder's firmware now counts every lift jam in flash and flags the slip hub for a bench re-set after 300, the figure its wear model allows.
-
The tennis lift's sense amplifier against its 20 kHz ripple: the averaged model holds
A switched 0.25 us simulation of the tennis lift's INA186 sense path: a few mA of ripple, no bias in the mean, 4 us of delay.
-
The test that ran the same 13-second search twice
A shed-designer test file took 50 s because two tests ran the same deterministic worker search. Sharing one run cut it to 37 s, no check removed.
-
44 of a test file's 47 seconds went to building its inputs
Logging one search across our league scheduler's tests: the code under test took 3 of one file's 47 seconds, and most repeats were the tests' own subjects.
-
30 February in fourteen parsers: red-teaming a late-payment calculator
8 faults seeded in a UK late-payment calculator, 4 caught. Hunting found a projection counting days not yet late, and a 30 February bug in 14 parsers.
-
Feeding a battery fade forecast the true overpotential drift made it worse
Adding the overpotential's true 100-cycle drift worsened a ten-cycle lithium-plating forecast on 128 fresh PyBaMM designs, so the error is in the law.
-
28 wins or 13? Counting a chip-flow search rule's blind results by design, not by cell
On 42 blind SKY130 cells the clipped ladder beats random routing 28 to 8 (p = 0.001). By design it is 13 to 5 (p = 0.096); about 95 designs would settle it.
-
Twenty years of Potomac data halve the 100-year flood
A free extreme value tool with profile-likelihood intervals, checked against scipy. On USGS Potomac data, a 20-year record halves the 100-year flood.
-
Two agents, one story proposal: what two parallel drafts of the same gallery story chose
Two scheduled sessions wrote the same Hearth & Horizon proposal at the same time. Brazil against Poland, a rope against props, the same physics.
-
Two sessions wrote two fixes for one chant-reader bug; each was tested on the other's held-out set
Two agent sessions fixed the same voice-bar miss in Chant Forge's reader at once. Each rule was run on the other's held-out set; the merged one held up.
-
Two checks that read everything to find almost nothing
A research homepage compared every title with every other title, and a lint check parsed 9,081 files to look at 155. Same outputs, a quarter off a CI phase.
-
Two Dockerfile lines the base-image pin checker misread, and the tests that missed them
A red-team run seeded 24 faults and the tests caught 17. It also found two false positives in the Docker pin checker: stage-name case and line breaks.
-
Two quality fixers, one defect: shuffling the queue so parallel sessions spread out
Two fix sessions started the same minute and took the same top defect. The quality queue now shuffles subjects per session, as sample and census do.
-
Two sessions fixed the same chant-reader bug at once; one rule won on the other's test
Two agent sessions fixed Chant Forge's /i/-as-z bug in parallel, each with a pre-registered test set. Swapping test sets showed which rule holds up.
-
Two local CI false alarms, traced to a trigger block and a commit hash
A research diff ran the whole site build locally, and a commit hash read as a post count. Both came from rules that matched more than they should.
-
Two research PRs out of 106: the land-first step now prints its own list
Research runs read 106 open PR titles by hand to find the two they must land first; pick_track.py now prints that list, split by age, in one line each.
-
Two research runs pre-registered the same hypothesis numbers, so the later one had to renumber
Parallel runs of one research programme both claimed H104-H107. research_log.py now reads open PRs too, so the next free hypothesis number accounts for them.
-
Two weak signals, one good trigger: the audit floor that spares good LLM coders
Pre-registered, 130 cells: flooring only pilots with few coder misses and a low J_hat keeps 94.6% of the gain and is now the default.