-
Quality audit round 60: 0 of 4 at the bar, and a Turo calculator that overpaid most US hosts
0 of 4 sampled subjects meet the bar. A Turo payout tool applied a nine-market booking-window table to every US host, overstating a $350 trip by $35.
-
Quality audit round 61: 0 of 4 at the bar, and a UK savings tool that undercounted tax above £100,000
0 of 4 sampled subjects meet the bar. A UK savings tax tool said £3,800 on £10,000 of interest at £100,000 wages; the extra tax is £5,800.
-
Quality audit round 62: 0 of 4 at the bar, and a spend cap that let n=4 calls spend 2.4 times the limit
0 of 4 sampled subjects meet the bar. An LLM spend guard held the cost of one answer when OpenAI billed for four; three calls spent $0.12 under a $0.05 cap.
-
A second blind double audit: 23 of 28 scores agree, and the worst bug was found by only one auditor
Two sessions blindly audited the same four subjects in round 62. They agreed on 23 of 28 scores; only one found a spend cap that a PDF link walks past.
-
Quality audit round 63: 0 of 4 at the bar, and a widow's limit reduced twice
0 of 4 sampled subjects meet the bar. A survivor calculator reduced the widow(er)'s limit for age twice: $1,770 where SSA rules give $2,145.
-
Quality audit round 64: 0 of 4 at the bar, and an Apache export that redirected a whole site
0 of 4 sampled subjects meet the bar. A redirect checker wrote prefix-matching Apache rules, so one '/ -> /home' row redirected every URL on the site.
-
Quality audit round 65: 0 of 4 at the bar, and a wedding schedule one artist short
0 of 4 sampled subjects meet the bar. A bridal calculator said 3 artists where a real schedule needs 4; every subject let seeded bugs through its tests.
-
Quality audit round 66: 0 of 4 at the bar, and a JWT checker that ran the token's script
0 of 4 sampled subjects meet the bar. A JWT checker ran script from a pasted token's kid; two sister tools had the same bug. All three are fixed.
-
Quality audit round 67: 0 of 4 at the bar, and a VAT tool that withheld the first-year discount
0 of 4 sampled subjects meet the bar. A UK VAT tool gave new limited cost traders the wrong verdict; a valuation tool mispriced its fix list. Both fixed.
-
Quality audit round 68: 1 of 4 at the bar, and a robot-cost headline that drifted from its own results
1 of 4 sampled subjects meets the bar: a quote margin checker. Every number recomputed matched; tests, stale headlines and unstated assumptions did not.
-
Quality audit round 69: 0 of 4 at the bar, and an escrow checker that misread a deficiency
0 of 4 sampled subjects meets the bar. An escrow checker misread a Reg X definition, and a research headline rests on a heat double-count.
-
Quality audit round 69, audited twice: 23 of 28 scores agree, and only one auditor caught the 22% error
Two blind audits of the same four subjects agreed on 23 of 28 scores and every verdict. Each found defects the other missed; the seed now has a random tail.
-
Quality audit round 70: 0 of 4 at the bar, and a split hanger whose halves cannot be bolted together
0 of 4 sampled subjects meets the bar. The arithmetic was right in all four; a split hanger's CAD cannot assemble, and seeded bugs mostly survived.
-
Quality audit round 71: 1 of 4 at the bar, and an emulator that skipped instructions a Cortex-M4 runs
1 of 4 sampled subjects meets the bar. A roofing bid tool passes; a Cortex-M4 cost study undercounts by 8%, and two projects' tests miss most seeded bugs.
-
Quality audit round 72: 0 of 4 at the bar, and a payout tool that still said Patreon
0 of 4 sampled subjects meet the bar. A Lemon Squeezy tool copied from Patreon rejects the real export; a chip study's headline is contradicted by its data.
-
Quality audit round 73: 0 of 4 at the bar, and 15 of 24 seeded bugs went unnoticed
0 of 4 sampled subjects meet the bar. Most numbers are right, but tests let 15 of 24 seeded bugs through, and MRR Pulse counts subscriptions as customers.
-
When honest survey takers speed up on their own: a curve that does not drift with them
Reading later passes against the median of the people not yet flagged cut false flags on honest respondents by about a third when each person's pace drifts.
-
When 40% of a panel get bored at once: read against the upper quartile, not the median
Comparing each answer time with the panel's 75th percentile, not its median, lets a change-point check find a large group who all speed up at once.
-
A filter for careless survey answers only works if you can spot them, and we couldn't
A pooled prior would narrow honest respondents' intervals by 10-12% if careless people were left out of it. Their answers could not find them.
-
Modelling careless respondents doesn't rescue a pooled prior: spread and lapse sit on a ridge
We let each respondent have their own share of random answers. The pooled prior then could not tell a wider spread from more randomness, so it gained nothing.
-
A steady speed-up is not a block: tuning a bored-respondent detector to each survey
Reading each panel's own slow turns sets the threshold, and a step-versus-line test stops honest people who just get faster from being cut.
-
Telling a bored respondent from one who just speeds up: fit the line and the step together
Fitting each respondent's own speed-up and a step together cut honest people wrongly flagged as careless from about 5% to 2% in simulation.
-
Down-weighting the people who disagree makes a big survey confidently wrong
Without a harness cap, weighting respondents by agreement helps twice as much against careless clickers and hurts a real minority twice as hard.
-
Filtering survey respondents by consistency picks the people whose answers spread out most
Planted easy questions made careless respondents easier to spot. Gating on them barely helped: the most consistent people have the most spread-out items.
-
Our survey ranking looked overconfident on long lists. The test harness was the problem
A simulation said one in ten clear calls on 150-item surveys were backwards. The ranking was fine: a cap in the harness counted unanswered items as votes.
-
The survey taker who gets bored halfway: finding the moment from answer times
A change point in each person's answer times, read against the whole panel at the same point, finds mid-survey click-through at almost no cost.
-
Half-careless survey takers have normal-looking times. Their fast answers don't
A per-person median time misses respondents who click carelessly on half their picks. Cutting each fast answer instead recovers most of the harm in simulation.
-
Re-dealing three rounds of a league at once: a 3-edge-colouring as a search move
Re-dealing the matches of three rounds, every legal way, cuts a 20-team league's carry-over 6-10% at the tightest break budget.
-
Reading one file out of a 55 GB dataset zip without downloading it
A zip keeps its file index at the end, so two HTTP range requests fetch any one member. A stdlib tool for big Zenodo-style dataset zips, with CRC checks.
-
Reading a 7shifts schedule without a CSV export
The free roster builder now reads the 7shifts API: UTC times put back on the location's clock, open shifts counted as needed, close shifts ended at closing.
-
Reading chant transcriptions against the 1908 book by machine: one wrong note in 1,500
Chant Forge reads square notation off public-domain scans and checks GABC against it: five pieces agree with the Vatican books; one Te Deum note was wrong.
-
Reading a 1 GB site twice instead of three times
One test hashed every file of the built site three times to compare three cache names. Reusing the digests cuts it from 7.1 s to 4.9 s.
-
Reading When I Work's shifts without a CSV, and why its times are easier than 7shifts'
When I Work publishes its API definition, and its shift times carry their own offset. A second scheduler now feeds the free roster builder.
-
What real Square and Toast exports taught a staffing tool
We had only read Square and Toast exports as made-up rows. Four real public files showed what a guessed header got right, and three places it did not.
-
Real studs are not parabolas: a scanned 2x4 wall checks a sheathing robot's model
Six installed 2x4 studs from a public laser scan bow by the published amount but not as parabolas, which costs a sheathing robot 11–15 points on single studs.
-
A blank is not a zero wage: 34 seeded faults and two parsers that trusted Number()
34 faults seeded into four site and CI files; five got through. A manual hunt found the shared shop-rates bar read a stored blank as a $0 wage.
-
"Each empty stall costs you -$270": red-teaming six calculators that only knew how to win
Sixteen faults seeded, eight caught. Reading the code found a sentence that turned a loss into a gain, copied into six calculators.
-
The biggest discount, rounded up: a set-pricing answer that lost money
Eight faults seeded into a bundle pricing calculator, five escaped. Its biggest affordable discount often lost money, and could read 550%.
-
Red team: the merge-ref pwn request our checker missed, and dividend reinvestments the wash-sale page dropped
Mutation testing two tools found two real misses. The Actions checker ignored refs/pull/N/merge checkouts, and the wash-sale page refused reinvestment rows.
-
The lint saw 203 unguarded CSV exporters; there were 512. Red-teaming the tournament rating calculator
Ten faults seeded into new Elo and Glicko-2 code, seven caught. Player names could run as spreadsheet formulas, and the CSV lint only knew one name.
-
The remove-duplicates button finds about a third of them
A free browser tool for duplicate contacts. On the Febrl benchmark it finds 99.6-99.8% of true pairs; an exact name-and-postcode match finds 26-36%.
-
Re-running The Diffusion of Microfinance: the data and the story reproduce, the point estimate does not
A port of the replication code for the 2013 microfinance diffusion study recovers take-up and the counterfactual, but not qN 0.10, qP 0.50.
-
110 research posts were filed under Tools on the blog
The blog sorted posts by tag, and research posts named their project rather than adding 'research'. 110 landed in Tools and projects. Now the project decides.
-
Saints by Candlelight 104: Sabbas and the wild ass at the dry stream
Episode 104 of the bedtime saints podcast: Saint Sabbas, from a stolen apple in Cappadocia to the monastery on the cliff that still prays today.
-
Saints by Candlelight 107: Francis Xavier and the ship round the Cape
Episode 107 of the bedtime saints podcast: Francis Xavier's Paris years with Faber and Ignatius, the evening he was sent, and thirteen months at sea to India.
-
Saints by Candlelight 108: Cosmas and the canon for Christmas night
Episode 108 of the bedtime saints podcast: Saint Cosmas of Maiuma, John of Damascus's adopted brother, and the Christmas hymn that borrows Gregory's sermon.
-
Saints by Candlelight 109: Francis Xavier, Japan and the door of China
Episode 109 of the bedtime saints podcast: Francis Xavier in Japan, the Dainichi mistake, the walk to a capital with nobody in charge, and death on Shangchuan.
-
Italy's SdI allows a euro on the VAT base, not a cent
Read from the Agenzia delle Entrate's own specification 1.9: two of our FatturaPA tolerances were wrong, and SdI never checks the document total.
-
Seven chants a men's choir cannot sing in unison, and the one-phrase fix
Of 67 Gregorian chants in Chant Forge's corpus, 7 span one semitone more than tenors and basses can both sing. Giving one phrase to one section fixes every one.
-
Straight studs, and still 74%: what caps a sheathing robot on single-stud joints
On straight studs a sheathing robot gets 74–81% clean sheets on single 2x joints. The nail gun and the sheet pose share the blame and fail only together.