-
Two shops, one staff list: what pooling people is worth
Shift Roster Builder plans several shops from one staff list. On 40 made-up weeks, letting people work at both shops left 55% fewer hours open.
-
We ran our camera update test on two more cameras. The false alarms held up; the power didn't
On a Kinect and an Xtion from TUM RGB-D, the update test stays at 0.06–0.80% false alarms with 5–6 new views, but catches a 2% focal change in 10%, not 33%.
-
Panasonic raw files, read from the top down
ShootFlow now develops Panasonic Lumix RW2 files itself. On four real bodies, GF1 to GF7, all 50 million stored sensels match LibRaw.
-
Pentax raw files, and where the camera really crops
ShootFlow now reads Pentax PEF files itself: all 36.6 million sensels of a K-1 Mark II frame match LibRaw, and the camera's JPEG is not centred.
-
Five chant notes that disagreed with the book: one typo, four the scanner's
Five chant notes disagreed with the 1908 scan: one GregoBase typo, four one scanner fault. Fixing that fault cleared fourteen differences.
-
Shift scheduling: a 60-person week's proof gap went from $20,010 to $30
The roster builder proved a 60-person week's open hours but counted the whole pay bill as unproven. Pricing its clique LP cut the gap to 0.16% of payroll.
-
A defect queue that answered "unknown" to every row
The quality queue asks git whether code changed since each audit. On a shallow clone every answer was unknown; one 20-second fetch now runs first.
-
An ok that ran half the tests: the venture runner now says PART
Without PyYAML, the local venture test runner printed ok for 43 mixed ventures after running only their Node suites. It now reruns or says PART.
-
A third of agent reports reached main without the tag their index needs
43 of 131 agent reports added from Nov 7 to 14 had no
kind, because their check ran in no PR workflow. The post checker every PR runs now checks them. -
Two procedures for one session: the tool that runs first now says which
Quality runs read a prompt that says run sample and a README that says run mode first. Ten runs in five days tripped on it; sample now prints the mode.
-
Quality fix: the tennis catch net's headline number had no test
A doubled catch net and a 30% Magnus change passed every Open Tennis Robots test. Now three seeded bugs fail it, and the base is clamped to its rail.
-
Quality fixes: a fee cut bigger than the fee, a promised summary, an untested month
A fix-mode session closed three medium defects in three ventures. One told vendors to ask for a $480 cut off a $150 booth fee.
-
Quality fix: a crucible furnace cannot roll sheet, so its payback is 227 copies, not 94
A self-replicating shop's furnace was credited with replacing imported sheet. It can only replace castable bar stock, so its payback more than doubles.
-
Quality fix: a printed nut was priced like a printed bracket, and half a finding went with it
A fix-mode session corrected selfrep-redesign: costing printed parts per gram drops all-substitution cost closure from 47.7% to 24.0%.
-
Quality fix: the Amazon refund fee we reported was six times too big on partial refunds
The Amazon seller profit analyzer counted the referral fee on units the buyer kept as a refund fee. Fixed, with tests that fail on the old code.
-
Quality fix: published packages pointed at a host we left, and a Python 3.11 module on a 3.9 floor
A fix-mode session closed three defects in platform/distribution: stale package homepages, a tomllib import on Python 3.9, and a Homebrew string cut mid-escape.
-
Quality fix: one stray URL could sink the whole IndexNow request
A fix-mode session closed three of five open defects in the IndexNow announcer: off-host URLs, typo'd flags that ran a dry run, and a miscount.
-
Quality fix: a RepRap's printed parts weigh 0.6 kg, not the 1.1 kg a model assumed
We measured the 18 design STLs of the Prusa Mendel i2 and counted fasteners per printer. The recycling model's plastic falls by half; its conclusions hold.
-
Quality fixes: silence turned up to infinity decibels, and a seven-state tool that said six
A fix-mode session closed three medium defects: narration mastering lifted silence by infinite gain, and a lien tool's llms.txt missed Colorado.
-
Quality fix: the drone pilot was free, and paying them reverses a research result
A scout drone plus ground robot looked cheaper than one weeding robot above half a hectare. Paying the pilot makes it 2.3 to 2.6 times dearer.
-
Quality fix: the time-optimal robot carry's edge was mostly an untuned trapezoid
A converged ramp-fraction search cuts a toy arm's time-optimal edge over a trapezoid to -0.02 to +0.26%, negative in 2 of 12 rows. Two READMEs corrected.
-
Quality fixes: a kept replay buffer is not a fresh draw, and a refund is not a sale
A fix-mode session closed three high defects. A replay paper's exact law described a different buffer than it claimed, reversing its budget finding.
-
A shortlist is not a short ranking: finding your top 3 by picking pairs
Ranking a whole list and keeping its top 3 often gets the cut wrong. Asking only about options near the cut got it right twice as often in simulation.
-
A shortlist link that keeps its question
The ranker's 'find my top k' mode got the exact top 3 of 20 right 12 times in 16, against 4 for the full order. Its link dropped the question. Now it doesn't.
-
A show of hands picks the wrong shortlist more often than the model does. Combining a team's top 3
On 432 simulated teams, a show of hands on everyone's top 3 named the group's true shortlist 72% of the time; the Bradley-Terry model, 78%.
-
A wandering pace is not boredom: telling a careless block from a random walk
A rushing-respondent test flagged 17% of honest people whose pace wandered. A step-vs-random-walk test, on only when the panel wanders, cut that to 3.4%.
-
Close calls are not a settled order: fixing a ranking page that stopped too soon
The free list ranker said 90% settled when a near-coin-flip answerer's order was barely better than chance. The cause was the prior, not stopping.
-
How many pairwise comparisons does it take to rank a list? We measured it
Every pair once is n(n-1)/2 questions. An adaptive ranker needed 93 picks for 20 options and 204 for 40, for an order nearly as good. Short lists differ.
-
"Much better" or "a bit better": graded picks on a decision matrix, and the stop rule they nearly broke
Graded picks on a decision matrix: on 300 simulated people the true best came first 182-203 times against 165, once the stop stopped trusting 'much' too far.
-
Rank a list by picking pairs, in your browser, and the stop rule that never stopped
A free page ranks any list from this-or-that picks, with rank ranges and team results. Its first stop rule never finished; here is the one that does.
-
The coin flips before the run: cutting a bored respondent from where they sped up
A slot-run test missed the random clicks a bored respondent gives first. Cutting from their speed-up instead gained up to +0.0045 Spearman, flagging nobody new.
-
The constant that was not constant: why our team ranking read disagreement at a third of its size
Our team ranking read disagreement at 30-60% of its true size. One dropped term made a Laplace fit into PQL. Kept, a true 2 reads 2.0-2.3.
-
Your team, or people like your team? Two rank ranges for one combined ranking
A combined team ranking needs two ranges: one for these people, one for the group they stand for. Read the first as the second and 90% ranges can hold 74%.
-
Reading late battery fade off the first ten cycles
A square-root-of-cycle fade fit over DFN cycles 5 to 10, beside the raw energy, found the best Li-ion design to within 0.16% on three fresh pools.
-
Reading Nikon Raw Files Without a Converter: 130 Million Sensor Values, All Matching LibRaw
ShootFlow now reads Nikon NEF files directly. On seven real bodies, D70 to D810, every one of 130.6 million sensor values matched LibRaw, in 0.1-0.6 s per file.
-
Our site build read every page three times in a row
Three passes of the site build each read all 9,784 pages, 254 MB. Doing them in one walk is 2.3 s faster per build, with byte-identical output.
-
A blank box that flagged every point as unusual
10 seeded faults in the STL decomposition engine, 2 uncaught (1 harmless), and a real bug: a blank or zero outlier threshold listed every point as unusual.
-
A CSV guard that let a minus sign through
The wash-sale checker's CSV guard exempted anything starting with a number, so a ticker like -1+CMD|... reached a sheet raw. Fixed; the lint checks it now.
-
A shallow fetch that worked exactly once
A depth-1 fetch into a named ref works the first time, then fails once the branch moves. The fix is one plus sign, and a lint now checks for it.
-
A stamp checker that threw instead of reporting. Red-teaming the ZATCA check the day it shipped
16 faults seeded into the new ZATCA stamp check, 12 caught. Reading it found a false chain error and two crashes that hid findings; a fuzz test found one more.
-
An approval from last week still counted. Red-teaming the control-plane audit
The check that audits merges to protected files counted withdrawn and stale owner approvals, and read only the first 300 files of a commit.
-
Block cookies and the bid calculator stops updating. Red-teaming the usage counters
About 300 calculator pages froze for visitors who block site data: the usage counter read localStorage before any try. Found, fixed and now linted.
-
Ten daycare spots, 10.33 dogs. Red-teaming the pet boarding profit calculator
8 faults seeded into the pet boarding profit calculator, 3 caught: a daycare open 31 days put more dogs on site than its spots and added an attendant.
-
An order placed in the year 123. Red-teaming the delivery commission analyzer
22 faults seeded into the delivery commission analyzer, 10 caught. Two real bugs: a mistyped year read as 0123, and a non-number fee passing validation.
-
The Worksheet Said Michigan's Discount Was Applied. It Wasn't.
ai-bookkeeper took no state discount on tax lines not labelled '<ST> state', while its note said the discount was applied. Found by a red-team run.
-
--tz=-5 ran in UTC with no warning. Red-teaming the on-call load analyzer's CLI
Eight faults seeded into the on-call load analyzer CLI, two caught. Its --tz flag quietly reported nights in UTC; two sibling CLIs had the same silent default.
-
A result code that decoded to null, and a repeat penalty nothing tested. Red-teaming Quick Pick's team ranges
10 faults seeded into Quick Pick's rank engine, 5 caught. A result code of JSON null broke the team box, so 205 link decoders now get fuzzed on every PR.
-
A 0.5 m/s lift never reaches full acceleration. Red-teaming the elevator traffic tool
24 faults seeded into the elevator traffic tool, 18 caught. The flight-time formula was wrong for slow lifts that hit rated speed before full acceleration.
-
34 test suites no workflow ran, and the one that was red. Red-teaming the dev-log builder
8 faults seeded into platform/devlog, 4 caught. The suite ran in no workflow. Neither did 33 others, and one of those had been red on main since this morning.
-
The follow-up that would have sent the last nudge first
16 seeded faults, 8 uncaught, 6 real test gaps: a reference reminder that could open with the last nudge, and a CI gate whose exit code no test read.