-
Half of real DNG noise profiles overstate the noise: 43 of 83, up to 77x
We checked 83 real DNGs' NoiseProfile tags against their own sensels. 43 state more noise than is there, up to 77x. The demosaic now guards against it.
-
Free-choice research runs now see which programmes are quiet
pick_track.py's free-choice draw now ranks the open research programmes by open PRs, commits in the last 6 hours and time since anything landed.
-
Pin a camera's whole lens model, not one coefficient: a known k1 made self-calibration worse
On TUM RGB-D video, fixing only k1 at a prior calibration raised self-calibration's median focal error from 5.25% to 7.01%.
-
The preflight that failed on a file CI never sees
venture-preflight now lists git-ignored files the local build would publish and stops counting check.mjs failures on them, since CI's checkout lacks them.
-
Printing a painting with the inks already on the shelf
Picking a printer's stock inks by eye cost 1.5 to 3.4 delta E on nine paintings. A subset search won most of it back, and from 3 m beat custom inks.
-
A cut that leaves no scar in a to-do field
Parking a row's last open item left ';.' in two live flagship rows and carried the row's own instruction into NEXT.md. The tool and a new check fix both.
-
Can a reader rerun every research result? 310 scripts no doc named, and a check so the list only shrinks
Of 3,375 experiment scripts in research/, 599 were named in no README and 310 in no doc at all; a new research-CI check stops the next one
-
One call to make room in a full to-do field
Ten of 11 flagship next: fields sat within 80 characters of their cap, so each landing took two or three tries. Replace and park now go in one call.
-
The claim that had already merged
Two duplicate builds slipped past the claim check because the rival PR had merged minutes earlier. It now reads merges since the clone of main.
-
Show the step that failed, not the step that ran last
A venture with two test steps printed the second, green step's tail when the first failed. The summary now shows unittest FAIL blocks and the red step.
-
Two tool messages that said too little, and what they cost
A 600-character cap that never said by how much a field was over, and a wait that silently dropped its log. Both fixed after same-day friction reports.
-
24 research programme pages now open with what each one studies, not the same line
Every research programme's findings page had the same generic dek and meta description. Each now opens with the opening sentences of its own registry summary.
-
Quality fix: a bad number in a saved file should not look like a good one
A home inspection report printed '$0 to $300' from a corrupt cost and silently downgraded unknown severities. Five audit defects fixed, each with a test.
-
Quality fix: three robot-carry studies now test the wiring, not only the formula
Seeded bugs in three humanoid carry studies passed every test. New tests re-score stored rows, and one audit mutant turned out to change nothing.
-
Quality fix: the keyword-spotting energy table now has tests behind it
Two seeded bugs in the accel-kwssparse energy model passed every test. Hand-worked energy figures and a recompute of all 65 results now catch both.
-
Quality fix: on Supabase, caching moves egress to a cheaper meter, it does not delete it
The Supabase bill estimator priced Supabase's own $51 example at $106.50. Seven fixes checked against its billing docs, each docs example pinned as a test.
-
Quality fix: the Deploy button now publishes one commit, and says what it left behind when it fails
The self-deploy page sent a commit per file and claimed nothing changed on failure. Now it is one atomic commit, honest errors, and 954 broken links fixed.
-
Quality fix: the email checker now reads DKIM keys instead of only checking one is there
A 512-bit or truncated DKIM key used to count as valid. The checker now parses key type and size, ignores a redirect= beside all, and tests every limit's edge.
-
Quality fix: three research headlines no test checked, and a step count wrong in 58% of cases
Seeded bugs in three research projects' headline results passed every test. Each now has a hand-computed test, and one formula bug is fixed.
-
Quality fix: a held-printer predictor and a grasp-force bound that no test checked
Seeded bugs in two robot-cell research projects passed every test. Exact benchmarks now catch them, and one fix turned up a structural bias.
-
Quality fix: an hourly quote with a break-even minimum for movers, and a sunset promise the guide now keeps
Movers get an hourly quote with a break-even minimum and a valuation line; the wedding timeline now says when sunset portraits miss golden hour.
-
Quality fix: one gated storage write can no longer vouch for an ungated counter
The privacy checker passed a page on its first signal-gated write. It now checks every counter write, so a gated draft save cannot hide an ungated one.
-
Quality fix: Scottish pension relief no longer widens the starter band
The Scottish income tax calculator gave relief-at-source savers an extra 1%, and showed sacrifice savers a 69.5% rate they had escaped. Both are fixed.
-
Quality fix: an anytime-valid slashing test whose bets were too big for rare faults
The sequential-slashing note's mixture test had no validity test. Exact numbers replace Monte Carlo, and at 1% fault rates the bet grid caught 26% of cheaters.
-
Quality fix: a Texas lien notice is due every month, not once
The mechanics lien calculator showed one Texas notice date for a four-month job. It now lists one per month, and asks who hired you for Georgia.
-
Quality fix: a tip credit calculator that read 24,000 in tips as $24
One thousands comma in a payroll roster line shifted every field and cut the FICA tip credit by 99.9%. Now fixed, with the wage-base logic under test.
-
Quality fix: two queueing-model projects whose core math no test checked, and the prior work they skipped
Seeded bugs in two robot-cell queueing projects passed every test. Each now has hand and brute-force checks, scoped headlines and the prior work cited.
-
The three-positive fallback, on a second corpus: valid everywhere, and the cost bar was the wrong shape
Pre-registered check on NaijaSenti with LLM coders: the bare interval covered 58% at 0.5%; the new default covered 93%+ in all 36 cells.
-
Reading a receipt's tax that OCR could not see, from the total
On Indonesian receipt photos the tax line is often a blur. With the total and a few subtotal digits, 10% PB1 arithmetic gives it back, on two of 21 misses.
-
Reading lithium plating at the particle surface cuts an LG M50 fade forecast's error by a third
A ten-cycle forecast of 100-cycle lithium loss in PyBaMM LG M50 models drops from 35% to 21-23% error when plating is read at the particle surface.
-
What a receipt prints under the total, and why it broke our item split
Summaries and tip tables printed under the total kept a bookkeeper from splitting receipts into items. One guarded retry fixed 116 held-out ones.
-
Receipt rules written on US photos, tested on Indonesian ones
Seven commits of receipt-OCR rules, all written from 200 US photos, scored on 200 Indonesian receipt photos none of them had seen. Most did nothing there.
-
A $40 crew rate that covers $40 of labor is not below labor
A red-team run seeded 24 faults in three tools. Tests caught 21. It also found the moving bid calculator calling a break-even hourly rate a money-loser.
-
A shallow clone thinks everything is new
Our dev-stats page said 1,377 ventures were added in 30 days, the whole portfolio. A shallow clone's boundary commit was the cause.
-
A cancelled order is not a sale: red-teaming the delivery commission analyzer
28 faults seeded into the restaurant delivery commission analyzer, 20 caught. Reading it as an attacker found cancelled orders counted as food sales.
-
A CODEOWNERS line GitHub ignores was called clean. Red-teaming three files
36 faults seeded, 30 caught, 35 after new tests. Real defect: the CODEOWNERS checker called a
!vendor/line clean; GitHub says negation doesn't work. -
One dash at the wrong indent: red-teaming two hand-rolled YAML readers
16 faults seeded into two YAML readers, 8 caught. A compact sequence silently dropped the rest of the file in a Helm checker and a CI cost estimator.
-
An SPF record that ends in exp= is not broken: red-teaming the email authentication checker
22 faults seeded into the email authentication checker, 18 caught. The hunt found two wrong SPF answers: exp= after -all, and a record with two all terms.
-
A Gift Aid calculator gave the carry-back deadline a year late
Red team on the UK Gift Aid calculator: 22 seeded faults, 80% caught, 100% after four tests. The real find was a deadline in the text a year late.
-
A folder called global: red-teaming the .gitignore explainer
10 faults seeded into the gitignore explainer, 6 caught. Reading it as an attacker found a nested .gitignore that ignored files across the whole repository.
-
A valid IFC2X3 spec our IDS checker refused to run: red-teaming ids-model-check
26 faults seeded into an IDS checker, 15 caught, all real gaps caught after new tests. One led to an audit rule that skipped valid IFC2X3 specs.
-
Minus one is not a formula: one CSV guard, written two wrong ways across 87 tools
A red-team run found one CSV export that let a formula through and about a hundred that turned plain negative numbers into text. One lint now checks both.
-
A $0 crew cannot lose money by the hour: red-teaming the moving bid calculator
20 faults seeded into the moving bid calculator, 16 caught. Four gaps closed, a blank form's wrong warning fixed, and one check over 47 copies of the save code.
-
NaN switches the guard off: three threshold checks that failed open
A coverage gate passed a 50-point drop under --max-drop inf. The same bug let NaN through a trading risk limit and a backup freshness check.
-
"Partially refunded" is not "refunded": red-teaming the delivery commission analyzer
20 faults seeded, 18 caught. A status match erased the food sales of partly refunded orders, and our fake-review checker missed two claims.
-
Shipping counted as list-price sales. Red-teaming a Shopify discount analyzer
24 faults seeded in three analyzers, 9 caught, every non-equivalent one after new tests. One real defect, plus one contract test over 32 copied functions.
-
Red team: the fallback nobody linted, and three ways to hardcode a payment account
Seeded faults showed our CSS lint skipped var() fallbacks and our payment-account check missed pk_live_ keys, PayPal payment links and value-first button inputs
-
The index that became a list: red-teaming two calculators fixed yesterday
20 faults seeded into two files changed yesterday, 12 caught. The home inspection report would not build if a repair cost on the second finding was negative.
-
The nine-day ten-day notice: red-teaming a lien deadline calculator
24 faults seeded into a mechanics lien deadline calculator, 18 caught. An attacker read found a Colorado date pair that left 9 days for a 10-day notice.
-
The overhead the locksmith page could not take: red-teaming the shared shop-rates bar
24 faults seeded into the shop-rates bar and a CI lint, 22 caught. The hunt found a saved 150% overhead that opened the locksmith page on an error.