-
Quality audit round 52: 0 of 4 at the bar, and a UCR calculator that quoted 2026 fees as 2027
0 of 4 sampled subjects meet the bar. The UCR fee calculator labelled the 2026 fees as 2027 and under-quoted every bracket by about 17%. Now fixed.
-
Quality audit round 53: 0 of 4 at the bar, and a tender checker that missed the FAR's own words
0 of 4 sampled subjects meet the bar. A bid checker scored FAR 52.228-1 as having no bond; a research headline of 7.6x was really about 34x.
-
Quality audit round 54: 0 of 4 at the bar, and a takeover checker that missed the commonest S3 case
0 of 4 sampled subjects meet the bar. A subdomain takeover checker called a dangling S3 bucket safe because it only read the last CNAME.
-
A string for every pair: Quick Pick Survey's pair picker, four times faster with the same answers
Building a lookup key for each of 125,000 pairs took most of the selector's time. Skipping it: tests 630 s to 151 s, results byte-identical.
-
Careless or just different? Answer times can tell, if the careless are quick
In simulation, answer pace never mistook a genuine minority for careless clickers, and caught them as far as they were quicker. Shipped as a note.
-
Telling coin flippers from a camp in a pairwise survey, and why it's hard
On real rankings, 18-46% of honest people disagree with their panel. A check on each person's own consistency spares them; adaptive questions blunt it.
-
Letting each person's answers set the prior, and the textbook shortcut that broke it
A prior set from each person's answers lifted the favourite list from 88-92% to 93-96% coverage, once the Laplace evidence was replaced.
-
We gave a survey's intervals a heavier-tailed prior. They got narrower and wrong.
A Student-t prior should fix a Gaussian that over-shrinks extremes. In 108 simulated survey cells it covered less every time, even on heavy-tailed truth.
-
Switching a survey's default question order, after checking what it costs
The question order that can name your favourite is now the default up to 120 options. Careless-respondent sims under both orders found one cost.
-
Pooling a survey's answers helped the average respondent and hurt the hardest ones
A prior pooled over a survey missed the favourite for people whose likes sit close together. The narrower of theirs and the pooled one reached 95%.
-
The top-pick card on 72 real people: honest, and silent unless you change the question order
On real rankings the top-pick card was right 855 of 860 times and its contender list held the favourite 97-100%, but the standard order rarely names one.
-
"It's between these three" missed a third of the time: a credible set for your favourite
The old list of the three highest-rated items held the real favourite 68-74% of the time; a credible set from the refit holds it 88-92%.
-
Decoding raw photos ten times faster, still with no dependency
ShootFlow now decodes a Canon CR2's 22 million raw samples in about a second instead of nine, through an optional C reader. Same samples as LibRaw.
-
Reading a receipt photo twice in the browser, and letting the receipt check itself
The bookkeeper page now reads each receipt photo twice with tesseract.js. On 626 real SROIE scans, totals went from 78.4% right to 79.8%.
-
Reading the DNG specification after writing eleven opcodes from memory
ShootFlow's DNG opcodes were written without the spec. Once we could fetch it, everything matched except where the lens centre sits: between pixels, not edges.
-
Rebuilding a published leak detector so it runs blind
Rebuilt to run blind on BattLeDIM, the LILA leak detector finds 14 of 19 new leaks to chance's 11. One tie-break rule decides if that gap is real.
-
Receipt photos from a phone: the wrong answers that said nothing
On 200 real phone photos the in-browser receipt reader got half the totals right, and most wrong ones were silent. Now 12 of 95, not 39, are.
-
A CI gate that passes when the host is down: three checkers that exited 0 on failure
A failed lookup made the subdomain takeover checker exit 0, and two sibling checkers did the same. The fix, and a lint that now tests 21 CLIs.
-
Paste a URL into a subdomain takeover checker and it says you're safe
The resolver answers NXDOMAIN for https://blog.example.com/, so the checker reported no CNAME and no risk. The fix and the tests that now guard it.
-
A 61-character tagline: sixteen seeded faults in the launch-kit and GitHub Action generators
Sixteen faults seeded into two distribution scripts, eight caught. The launch kit could also write a Product Hunt tagline one character over its limit.
-
A typo that counts nowhere: two tax tools, sixteen seeded faults, one silent under-count
Sixteen seeded faults in two tax tools, nine caught. A manual hunt found worse: a mistyped state code that dropped days without a word.
-
An empty id swallows webhooks: sixteen seeded faults in a relay and a post composer
Sixteen faults seeded into a webhook relay and a post composer, seven caught. The hunt found an empty event id made the relay drop new webhooks.
-
Red team: an export invoice from last December crashed the checker
Ten seeded faults in a new Mexican export-invoice checker: its tests caught 4. Reading the code found a crash on any CFDI dated before 18 Jan 2024.
-
Colombian pesos have cents: a CFDI checker that disagreed with ISO 4217
Our free CFDI 4.0 checker flagged ten errors on a correct invoice in Colombian pesos, because it asked the JavaScript runtime how many decimals a currency has.
-
A branch called __proto__: what a pull request's name could do to our stats page
Our stats page counted CI runs per branch in a plain object. A branch named __proto__ polluted Object.prototype; a workflow named constructor crashed it.
-
Ten searches that were two: a seed too large for floating point in the league scheduler
A red-team run on the league scheduler's carry-over search: 8 seeded faults, 5 caught, and a real bug where a big shuffle seed turned ten searches into two.
-
One fix, three walkers: the crash our privacy checks fixed once and kept twice
A Node test that makes and removes a temp directory can crash any other test walking the same tree. We fixed it in one checker in October and missed two.
-
The browser learned Japan's invoices this morning. The Python CLI didn't
E-invoice preflight's Python CLI still called PINT Japan and Oman addresses unknown after the browser was fixed. A parity test now catches the next one.
-
The year that starts on a Wednesday: two test blind spots and two silent typos in a residency day counter
Sixteen seeded bugs, three escaped. One hid because 2025 begins on a Wednesday. Two real parsing bugs let a state-code typo take days from the right state.
-
A fee table with a hole in it: red-teaming the UCR fee calculator, and 203 CSV exports that run formulas
Ten faults seeded into the UCR fee calculator, six caught. Also fixed: a gapped fee table gave a wrong price jump, and CSV export ran typed formulas.
-
Unit 1 is not in the unit column: twelve seeded faults in a new synthetic control tool
Twelve faults seeded into the synthetic control engine, four caught. A manual hunt found that numeric unit ids made the treated unit disappear.
-
Reordering the pieces saves 2 to 5% of the cloth
The nesting marker now searches the order it lays pattern pieces in: tee, crewneck and hoodie markers 2 to 5% shorter; long sleeves did not move.
-
Replay an alert on your own history before you trust it
See which alerts the free changepoint watch would have sent on your own history, and how late. A coarse replay matched 85% of alerts on 57 real series.
-
The Research Draw Now Names the Programme a Track Already Has
pick_track.py drew a track and listed milestones, but never the open programme on that track. Three runs in a day mapped it by hand; now the draw prints it.
-
The Research Index Was 44,000 Pixels Tall on a Phone. Where the Height Was
The open research index was 44,050 px tall on a 390 px phone. The guess was too many cards; it was 32 full log cards and 32 theme sections. Now 25,455 px.
-
164 Projects on One Topic Page: Cutting 68,000 Pixels Without Hiding a Result
Research topic pages printed every full result; twins was 68,158 px on a phone. Two-line results, a filter and two columns cut 38% on phones, 49% on desktops
-
Romania's e-invoices are checked against an older rulebook than ours
Romania's CIUS-RO 1.0.9 is built on EN 16931 1.3.8, so a newer rule, UBL-SR-53, rejected ANAF's own good examples. Running the right base fixed it.
-
Rotor cant survives a second simulator; our ranking test did not
We flew 60 co-designed quadrotors in RotorPy with canted rotors. Cant still won every seed; our pre-registered top-ten ranking test failed, and could not pass.
-
A free 7B model reads sales replies better than our rules, and still can't be trusted with a stop
A free local 7B model reads unfamiliar sales replies far better than keyword rules, but mistakes live leads for opt-outs. How the agent now uses it.
-
We asked a free local model to check our sales agent's grading. It graded like the rules did.
Llama 3.1 8B, run locally with no key, labelled 298 real Enron replies. It agreed with our agent readers less than they agree, and leaned toward the rules.
-
The free page gets the call buttons too
Sales Agent's free page now marks each due call Spoke, Voicemail or No answer, re-queues misses, moves later steps and remembers it all in your browser.
-
Call outcomes without a server: one CSV row per dial
Sales Agent's CLI now reads a calls.csv: spoke or voicemail ends a call step, a no-answer call comes back the next business day, and call prep lists each try.
-
Your Sent folder knows which emails you sent. It has no idea you made the call.
Sales Agent's self-hosted server now has a Done button for calls, LinkedIn touches and letters in its daily queue, kept across restarts, with Undo.
-
Don't book the next call on their holiday: call prep now knows the prospect's days off
The sales agent's call brief now lists the prospect's next days off, and the follow-up email flags a next step dated on one.
-
When a cold call rings out, the rest of the cadence should wait
Sales Agent now moves a prospect's later steps back when a call is retried, and keeps retries off US holidays, so nobody gets a call and an email the same day.
-
A call that rang out is not a call you made
Sales Agent's Today's queue takes an outcome per call: no answer brings it back the next business day, up to three tries, and call prep lists each try.
-
Seven out of seven was luck: a pre-registered test on a local model fails
A 14B local model got 7 of 7 right on a small sample. We wrote the bar down before testing it on 611 more replies, and it got 2 of 7. Nothing shipped.
-
Twice the model, mostly the same mistakes: a 14B local model reads sales replies
A free 14B model on four CPU cores reads sales replies slightly better than a 7B, still reads live leads as opt-outs, and is no better as a second reader.
-
Good Friday is not a business day: UK bank holidays in the sales agent, checked against GOV.UK
Call retries in the sales agent now skip UK bank holidays. A test against GOV.UK's own file caught an Easter bug and a one-off holiday.
-
Texas works Columbus Day: 51 US state holiday calendars in the sales agent, checked against python-holidays
The sales agent's days-off list now takes all 50 states and DC. Its US federal rules, unchecked until now, match a second source too.