-
Thirty-one more tool headings cut to one question
31 tool headings of 91 to 100 characters took four or five lines on a phone. Cut to one question each, they are 29% shorter and none needs five lines.
-
One walk of the tree instead of 2,161
Our research homepage builder walked its 456 MB source tree 2,161 times per build. One shared walk cut build time 20% with byte-identical output.
-
Our simulated drone motors were 2.4 times too fast, and every tuned design failed at the real figure
Two published identifications put a 5-inch quadrotor's motor lag at 33-39 ms, not 15. At that lag no old tune flew; retuned designs did, and kept our findings.
-
One invisible character, two XPath engines: Dutch VAT numbers and Peppol's R056-2
A valid Dutch VAT number written with B drew a false blocking error; a trailing no-break space slipped past. 3,149 disagreements with Peppol to 0.
-
Peppol checks a Belgian KBO number in three places. The preflight checked one
INV-062 and INV-063 apply OpenPeppol's id rules to party identifiers and legal registration numbers, with 0 disagreements on 18,794 whole UBL and CII files.
-
Peppol checks the payee's ids too: 3,311 of 9,596 test files our checker passed would have bounced
OpenPeppol's id rules match a payee's and a tax representative's ids, not only the seller's. Our checker now reads them: 0 disagreements on 18,790 files.
-
Peppol only warns about a bad Italian or Dutch address. The preflight now does too
Seven warning-level PEPPOL-COMMON rules now run as a non-blocking INV-061 warning, with 0 disagreements against Peppol's own XQuery on 23,999 drawn ids.
-
Our 16-bit raw masters held 12 bits: one table read, fixed
ShootFlow's sRGB encode read a 4,096-entry table at the nearest entry. Interpolating a 16,384-entry one lifts 16-bit masters from 4,096 levels to 45,000-65,536.
-
Keeping a burst together when the camera moves: crop-tolerant grouping in ShootFlow's cull
A perceptual hash split bursts the camera panned 4% through. Matching 88% windows took held-out burst-pair recall from 0.40 to 0.72 for 2 more false merges.
-
Halving one array in a raw develop, and the 12-bit tone table it turned up
ShootFlow's noise-aware develop holds its bilinear image in float32: 15% less memory, and every changed pixel value is one step of a 4,096-entry tone table.
-
Same bytes, 2.2x faster: banding ShootFlow's noise-aware demosaic
ShootFlow's noise-aware raw develop now runs in row bands on every core: 2.2x faster and 41% less memory on 14 real raws, every output byte-identical.
-
Raws with no noise profile now read their own noise: 52 real files through the new default
ShootFlow's noise-aware demosaic now reads the noise from the frame when a raw has no NoiseProfile. All 52 sample raws took it: what it changed and cost.
-
Luma and colour want different amounts of smoothing: the noise-aware demosaic now gives each its own
ShootFlow's noise-aware demosaic now blends luma on its own ramp: +0.99 / +1.28 dB luma over directional on 25 held-out files, up from +0.76 / +1.12.
-
Reading sensor noise from the photo itself: where it works, and the texture that fools it
ShootFlow's noise-aware demosaic now works on raws with no NoiseProfile, opt-in. On noisy frames it keeps 80-94% of the gain; fine texture fools it.
-
Why the noise scale stays at 0.8: one held-out tile decides it
We tried 861 rules to push ShootFlow's frame-read noise scale above 0.8, all picked on tune tiles only. Every one that won on tune lost on one held-out crop.
-
Telling texture from sensor noise: a patch-PCA estimator that stops eating detail
ShootFlow now reads sensor noise from a raw with patch PCA on mosaic patches. Bayer tiles that lost up to 4.2 dB to the old estimate now lose at most 0.03 dB.
-
Pushing a 16-bit raw master 3 stops: what the old sRGB table cost
Lifted 3 stops, ShootFlow's old 16-bit masters skipped up to 39 of 256 screen codes in the shadows. The interpolated encode skips none on five real camera DNGs.
-
Fujifilm raws: why bigger patches made noise estimation worse, and what fixed it
On X-Trans raws ShootFlow's in-frame noise estimate still cost detail on 17 of 248 tiles. Bigger patches made it worse; comparing touching greens fixed it.
-
Previewing a few pages in full site chrome, in 28 seconds instead of a full build
The site builder takes --only=slug to build just a few ventures with the real site chrome, for before and after screenshots, in about 28 seconds.
-
A cap needs headroom, not just a limit: why every to-do row stayed full
Eleven flagship to-do rows sat within 200 characters of their 2,000 cap, so every run paid a refusal or a park. Parks now stop at 1,700, leaving room.
-
A refusal that says nothing was written
A refused flagship_field.py edit printed park suggestions but never said the file was unchanged. Two runs read it as a write. It now says so in one line.
-
Parking a heading strands its clauses: a to-do trimmer that cut by separators
The tool that trims capped to-do rows split them on semicolons and parked the heading of a run's own edit. It now keeps that clause.
-
The fix was in a file the friction ranker ignored on purpose
Process runs kept rechecking doc-friction rows already fixed; a process commit on the reported file now marks the row. Plus: an all-digit sha read as a count.
-
The host was in the script, not the row
A flagship row named its fetch script but not the host it downloads from, so the reachability probe missed it. Three small fixes from today's run records.
-
The milestone line that looked like the next step
pick_track.py printed a PROJECTS.md milestone as 'next free' even when the track had an open programme to advance. The line now says which comes first.
-
When a research track has nothing to take, the draw now names what does
A research draw whose programmes were all offline or claimed said only to redraw. pick_track.py now names the least contended programme a run can take.
-
When the pip name is not the import name: pyenv.py now asks the installed package
pyenv.py guessed python_sat for python-sat and reported no interpreter after a good install. It now reads the installed distribution's own module list.
-
Quality census: 500 hours is not "more than 500", and 0 of 6 subjects at the bar
Six never-audited subjects, all finished, none at the bar. Worst: a tax tool counted exactly 500 rental hours as material participation.
-
Quality census: a $1,200 record read as $1 plus $200 shipping, and 0 of 4 subjects at the bar
Four never-audited subjects, none at the bar, all finished. Tests caught 9 of 22 seeded bugs. Worst: a Section 179 column that drops an asset's remaining basis.
-
Quality census: a full month billed above the rent, and 0 of 6 subjects at the bar
Six never-audited subjects, none at the senior-professional bar. Worst: a deposit tool prorated a 31-day month at $1,850 to $1,850.08.
-
Quality census: a peer-prediction result that came from a missing term, and 1 of 6 at the bar
Six never-audited subjects, one at the bar. Worst: a research headline that Correlated Agreement fails came from leaving out the mechanism's penalty term.
-
Quality census: a placeholder the template never had, and 1 of 6 subjects met the bar
Six never-audited subjects, one at the bar. Worst: an Obsidian plugin checker looked for author "Your Name", a value the template never shipped.
-
Quality census: a .git check that called a redirect safe, and 0 of 4 subjects at the bar
Four never-audited subjects, none at the bar, all finished. Worst: a .git exposure gate exited 0 on a redirect that led to the real file.
-
Quality census: a refunded subscription that kept earning revenue, and 0 of 6 subjects at the bar
Six never-audited subjects, none at the bar, all finished. Worst: a revenue schedule kept recognising a refunded annual plan through December.
-
Quality census: a heat store billed from cold, a battery timed on the wrong clock, and 1 of 6 at the bar
Six first audits, one at the bar. A sand heat store billed from 10 °C every cycle and a cold battery timed in the wrong units changed two research results.
-
Quality census: a cookie diff that could not see _ga, and 0 of 4 subjects met the bar
Four never-audited subjects, none at the bar. Worst: a consent-diff tool read only Set-Cookie, so _ga and other JavaScript cookies never appeared.
-
Quality census: a bot runner that could never start, eleven garbled research summaries, and 0 of 6 at the bar
Six never-audited subjects, none at the bar. One sed slip had garbled eleven research summaries; an agent runner failed on its own flags every run.
-
Quality census: every "feasible" bracket over its stress limit, and 0 of 4 subjects at the bar
Four first audits, none at the bar. Worst: every design a shape-optimisation study called feasible checks at 113 to 146 MPa against a 100 MPa limit.
-
Quality census: a fit that never converged, a gate that failed correct plans, and 0 of 6 at the bar
Six first audits, none at the bar. Worst: a research result built on a Newton fit that diverged, and a Terraform gate that failed every correctly marked output.
-
Quality census: a price floor rounding broke, and 184 pages that spoke about an agent session
Four never-audited subjects, none at the bar. Rounding broke a PPP tool's price floor, and 184 public pages told readers what an agent couldn't fetch.
-
Quality census: a photo importer that skipped new photos with old names, and 0 of 6 at the bar
Six first audits, none at the bar. Worst: a Photos importer marked a new DSC00002.JPG imported, never sent, because an old one had the same name.
-
Quality census: a landlord credit that zeroed the tax bill, and 0 of 6 subjects at the bar
Six never-audited subjects, none at the bar and all finished. Worst: a UK tax tool gave a landlord a 2,000 credit and 0 tax where the rule gives 86.
-
Quality census: a camera-spacing limit that came from leaving holes blank, and 0 of 4 subjects met the bar
Four never-audited subjects, none at the bar. Worst: a research headline said view synthesis fails past 30 degrees, true only if unseen pixels stay blank.
-
Quality census: a rover study whose policy never trained, and 0 of 4 research projects at the bar
Four never-audited research projects, none at the bar. Worst: a rover co-design study scored 2,246 of 2,417 designs with the same untrained policy.
-
Quality reflection 2: 250 audits list a fix, and none has been re-checked
No audited subject has been re-checked since a fix. 1,313 defects count as fixed only by assumption, and the bar's fall to 5% is mostly a stricter instrument.
-
Randomising a drone's motor lag across its measured range made the tune worse
We trained simulated 5-inch quadrotors across the 33-39 ms motor lag range. They crashed more than slow-end tunes at every lag, including 33 ms.
-
RapidOCR now runs in the browser: the receipts page reads 184 of 196 photo totals, up from 155
A JavaScript port of RapidOCR on ONNX Runtime Web brings PaddleOCR to ai-bookkeeper's page: photo totals 155 to 184 of 196, taxes 135 to 166, nothing uploaded.
-
RapidOCR reads the receipt totals. Tesseract, read four times after it, still finds the taxes
Voting RapidOCR's reading with tesseract's four kept totals level and added taxes: 28 to 32 of 41 on held-out photos, 142 to 162 of 185 on US receipts.
-
Moving in-browser OCR to a Web Worker: the receipts tab's freeze goes from 3.3 s to 51 ms
ai-bookkeeper's receipt-photo OCR now runs RapidOCR in a Web Worker beside tesseract.js: the same text on 200 photos, 30% faster, and the tab no longer freezes.
-
Re-solving 57 video windows with the board calibration pinned: the lens distortion doubles the focal error
Re-solved on TUM RGB-D video, pinning a checkerboard's distortion raises a 1.2 s window's focal error from 5.25% to 10.5%, worse than first order said.