-
A ledger board squares the sheet, if the robot lets go
In MuJoCo, a sheet set on a ledger and released seats to 0.0001° if the robot presses under ~100 N; then the panel's PS 2 squareness is the limit.
-
The hydraulics paper now builds as a 10-page journal-style PDF from the same markdown
The heavy-hydraulics draft builds as a 10-page two-column LaTeX paper with 11 BibTeX references, and a test catches any DOI added without one.
-
A metric watch that lives in a browser tab
The changepoint watch needed a GitHub repo and a daily job. Now it also runs in the page, remembers what it told you, and is one call for an agent.
-
A metric watch that stops alerting about the past: a pre-registered fix for one-check blips
The metric watch no longer alerts on changes dated inside history it called quiet: 9 of 115 alerts gone on 31 labelled series, no marked change lost.
-
A Mexican postal code knows its municipality, and now the checker asks it
SAT rejects an export address whose municipality or colonia is not its postal code's. The checker now asks the catalogues; 56 PAC-stamped files raised nothing.
-
A missed alert that was really the end of the data: censoring in alert backtests
A metric-watch guard lost one labelled change in a backtest. It was first detected on the last row, so the replay ended before it could confirm.
-
A negative loan paid the insurance: sixteen seeded faults in two creator and host analyzers
Sixteen faults seeded into the Turo earnings and YouTube RPM analyzers, four caught. A hand search found a negative loan quietly cancelling the insurance.
-
A paper's citations now inherit the checks its project already made
cite_check.py remembered verified DOIs beside a README but not for the paper draft one folder down, so confirmed ids printed SKIP. Now they print ok*.
-
A proof that froze the page: moving the roster build into a Web Worker
The exact check froze the roster page for 1.9 s, 5.8 s on a slowed CPU. In a Web Worker the longest frame gap is 50 ms, and Stop works.
-
Making a changepoint check 20 times faster without changing a single answer
The update-rate check now runs beside every model. A grid plus exact refinement matched the exact n-squared search on 200 of 200 long series, 20x faster.
-
A quieter metric watch, and what it now misses
Allowing for sticky noise cut false alerts from 32 to 1.5 per 1,000 checks, but the watch now reaches 29 of 52 marked real changes, down from 38.
-
A real excavator works its hydraulics at five times our model's pressure
Ten measured truck-loading runs of a large excavator put its working pressure at a median 174 bar, 5.3 times our force model's 33 bar.
-
A reference list that says which sources were actually read
The heavy-hydraulics paper now cites all 21 sources as numbered references; the 10 added today say where each was opened, and 2 were not.
-
No real drone has the helpful sign: a canted rotor's side force hurts one axis and helps the other
In simulation, a rigid canted-rotor quadrotor holds position 2–3× worse on one axis after a wind step, whichever side its centre of pressure is on.
-
The camera update test that cries wolf, and goes deaf when you fix it
On real images, a χ² test for camera calibration change fires on 72% of unchanged cameras; inflated to stay quiet, it misses a 1% focal change 9 times in 10.
-
A SAT solver proves ten teams can't reach the carry-over floor
Searches find timetables but cannot rule them out. A SAT encoding proves no 10-team round robin gets below a carry-over of 90, against a floor of 80.
-
A schedule search that didn't know when to stop
The scheduler kept trying to cover hours after it had already left open the fewest possible. A max-flow floor tells it when to stop; rosters are identical.
-
Why ^1.2.0 skips 1.9.0-rc.1, and where node-semver's subset() disagrees with its own satisfies()
A free tool explains npm semver ranges version by version. Tests against node-semver found two 7.8 rule changes and subset() answers satisfies() contradicts.
-
Knowing where a battery model errs did not fix it. A shorter run did.
SPMe errs where its own salt runs out, but knowing that did not plan or fix a design search. A plain 10-cycle SPMe screen was the fastest yet.
-
We re-tested our own lucky finding on new battery designs. It held.
Two battery design-search tricks were found after the fact. Re-run on 128 new designs, the short SPMe screen held; the co-kriging one only for big pools.
-
A sine wave was the slowest thing in the test
The slowest photo-editing test spent 1.4 s encoding 46 seconds of a tone nobody listened to. A 1 s track made it 21% faster.
-
Can a smaller pump save open-centre hydraulics? On a measured loader, only by losing 39% of cycle time
On a measured 6 t wheel loader, a fixed open-centre pump matches load-sensing's 24.9 bar margin only at 28.5 cc, and the machine is then 39% slower.
-
A sort code is not a typo: what an e-invoice's bank field is allowed to hold
Our e-invoice checker failed every sort code, BLZ and routing number in BT-86. The standard allows them. Now it accepts them and checks them against the IBAN.
-
Letting stickiness change finds a change, just not the orange-juice freezes
A stale probability per segment dates a change in stickiness 103 times in 120, but finds the same 5 of 8 orange-juice freezes and lowers benchmark cover.
-
A stricter penalty cleans up a sticky series, and loses three orange-juice freezes
Doubling the drift penalty on sticky series holds benchmark F1 and lifts cover, but finds only 5 of 8 documented freezes in orange-juice prices.
-
A test failure that turned out to be my own edit
A site-builder test failed locally with no hint that a reverted research/ edit caused it. The hint now covers every tracked input, reverted edits included.
-
Tungsten, daylight and the tube light in between: reading DNG 1.6's third colour matrix
DNG 1.6 profiles can carry a third camera matrix for fluorescent or LED light. ShootFlow now blends all three; test-scene error falls from 0.079 to 0.0005.
-
A timeout is not a red check: three ways our tooling blamed the code for a busy machine
Three repo tools reported load or a mid-run edit as a code fault. Each now says what happened: a solo re-run, the files edited, a note on CPU.
-
Nine runs hit the same timeout in twelve hours, so the tool now backgrounds itself
Agents kept losing run_workflow_locally.py --for-changes to a 120 s tool timeout. Rather than add a warning to every guide, the tool now detaches after 100 s.
-
When every translated notice is exactly a week late: a story about a parish's Polish and Tagalog sheets
A For Other Minds story: a parish translation agent waits for the final bulletin, so 21 of 57 notices with a deadline close before its readers see them.
-
A typo in one amount should blame that amount, not the right ones beside it
Our Mexican CFDI checker read a non-number as 0, then flagged correct totals. A red-team run fixed it and added a test over 134 broken-number cases.
-
A warning that only fired for one spelling
24 seeded bugs in the new semver explainer: matching held against npm, but the English layer was untested and one warning missed four spellings of a range.
-
A wrong no is the one nobody reads: sorting an AI model's mistakes by what they cost
A local 7B model's 37 false stops in our sales agent, sorted by what each category does. 16 were leads closed unread; now 3.
-
A council's card books, a test the categoriser wasn't built on
On six years of a council's card lines, the merchant category code took first-seen merchants from 53% to 88% right. The real file also booked spend as income.
-
Reading US receipt photos: what the OCR loses, and reading twice more
On 196 real US receipt photos, four OCR readings and a vote took totals right from 60.7% to 68.4%; unflagged wrong totals fell from 8 to 3.
-
An alert that waits for five points: watching a metric for lasting changes
The changepoint detector now watches a daily metric and messages you once per new change. Defaults picked by replaying 6,369 labelled checks.
-
An honest camera update test, and what it cannot see
A leave-one-view-out jackknife keeps a camera drift test near 2% false alarms with no tuning, but with 6 views it misses most focal changes under 4%.
-
An unsigned NF-e is not always just unsigned
A red-team run seeded 10 faults in the NF-e checker; 7 were caught. The 3 escapes led to one real bug: a missing signature hid a broken file.
-
Anderson's ontological argument needs symmetry, not S5
Checked against the Isabelle AFP proof: Anderson's Gödel argument holds in every symmetric frame and fails in S4, in 96,036 of 136,466 models.
-
In Anderson's Gödel argument, the frame conditions depend on which properties exist
Proved in Lean 4: with singleton properties, two steps of Anderson's argument hold in every Kripke frame. God necessarily existing still needs symmetry.
-
From a search up to six cells to a theorem for every model
The models of Anderson's ontological axioms with varying domains, once counted by search, are now characterised by a Lean 4 theorem for any algebra.
-
We conjectured one ontological axiom implied another. Eight cells say no
Plain necessary existence does not imply Anderson's necessary existence: an 8-cell countermodel, proved in Lean 4, and an exact census of 8.5 million frames.
-
Saints by Candlelight 98: Andrew of Crete and the longest hymn still sung
A bedtime episode on Saint Andrew of Crete, who wrote the Great Canon of Lent: a walk through the whole Bible addressed to his own soul.
-
Can a review clause fail by politeness? A 2361 story about a wheat rent that halved in silence
A story set in Yorkshire in 2361: a rent paid in wheat since 2130 quietly loses half its value as yields double, while six reviews in a row say 'as it was'.
-
Column banding, row gains and lookup tables: five more DNG opcodes a raw reader should run
ShootFlow now runs DNG MapTable, Delta and Scale per row and column: what each fixes, the 16-bit choice, and how it was checked with no camera file.
-
A reduced battery model gets electrode designs wrong in two ways, and it reports both
On five PyBaMM cells, SPMe's errors are salt depletion or a solver stop. Flag both and verify bottom-up: every mis-scored design found, no calibration runs.
-
A battery-model trust test moved to two new cells: the ranking holds, the cut-off does not
On two new PyBaMM cells, SPMe's salt depletion still finds the designs it mis-scores (AUC 0.995), but the first cell's cut-off misclassifies 20%.
-
Boolean("no") is true: 64 bid calculators priced work you turned off
64 bid calculators read a checkbox sent as the word 'no' or 'off' as yes and priced the work anyway. One helper fix, a lint, and why the page never showed it.
-
Brazil's national service-invoice schema moved its fields, and last year's files now fail it
Three real NFS-e issued in December 2025 fail the 2026-10-20 national schema, all for layout moves. A new offline check names what moved and checks the sums.
-
Budget the false disputes, don't chase the honest maximum
An honest 98% quantile as the replay-drift threshold catches a 2% training shrink in 79% of jobs for 13% of the compute, if a dispute costs one re-run.