-
Hearth & Horizon: three turns on a bollard, and a rope that changed
A new Hearth & Horizon story: three turns of polyester hold a 3-tonne boat with 4 kg; three turns of slippery HMPE rope need 100 kg.
-
Which battery designs can you throw away before simulating them?
Dropping electrode designs whose closed-form salt number is below -0.2 kept the best design on three fresh PyBaMM pools and cut the screen's cost by 21 to 52%.
-
When 12,50 was read as 1,250
33 CSV tools took the decimal format from the delimiter, so a tab-delimited European file read 12,50 as 1,250. They now read the amounts themselves.
-
For Other Minds: a pantry's wait sign, the people who left, and the people who never took a number
A new For Other Minds story: a food pantry agent's 12-minute wait sign counted only people who were seen, and its honest fix sent people home unlogged.
-
Twelve teams at 132: the wall at 136 was a plateau
Two C searches stopped at 136 and called it a wall. A third, walking plateaus, found a 12-team round robin at 132; the proved least is 124.
-
Twelve teams cannot reach 122: proving a carry-over lower bound with one matching per round
An exact bound for the rest of a round and a round-1 symmetry cut made an exhaustive search 11x faster; no 12-team round robin has carry-over below 124.
-
Two cycles at once: what a daily-only decomposition calls an anomaly
The free STL tool now runs MSTL for several seasonal periods, matched to statsmodels. On hourly grid load, the one-cycle fit's 'outliers' were weekends.
-
Two cheap cycles can't see the plating fade
A two-cycle SPMe plating predictor on 2,324 Li-ion designs: the plating rate doesn't rank 100-cycle fade, and SPMe fails on 20% of the designs.
-
Two errors cancelled in our camera update test; with an honest reference it catches 19%, not 54%
A MAD jackknife against χ² seemed to catch a 2% focal change in 54% of cases. Its cut-off is too lenient; a model cut-off gives 19% at 0.2–0.5% false alarms.
-
Two gear-pump makers, opposite answers, one number: 0.83
Casappa's curves say load-sensing wins, Danfoss's say a smaller fixed pump does. On one efficiency axis both flip at about 0.83 overall at 150 bar.
-
Two million regex calls to find zero bugs: a lint nine times faster
A calendar-date lint ran six regexes on every line of 5,704 files. A fixed-text check every match must pass cut it from 3.5 s to 0.4 s, same output.
-
Two GPIO pins and a cable the clash check never saw
Routing a DIY tennis feeder's cables found a pin clash that would have stopped two motors, and a CAD fuse that hid a cable from the clash checks.
-
Missouri and New York pay you to file sales tax on time. The worksheet now takes it off.
The sales tax worksheet now applies Missouri's 2% allowance and New York's 5%-up-to-$200 credit, read on the states' own pages.
-
Two sensors agree on noise too: a pre-registered test of 'alert when two columns agree'
On a real 8-sensor pump, waiting for two sensors to agree cut false alerts from 4.2 to 2.7 per 100 checks, still over budget, and missed 9 of 33 faults.
-
Two thirds of a loop was skipping: the ad budget optimiser's test drops from 21 s to 6 s
The ad spend optimiser's exact budget split mostly checked totals no channels could reach. Bounding the loop made a test 3.4x faster, same answers.
-
Auditing outsourced training: when you can't deter cheating everywhere, deter it somewhere
Skipping 20 training steps costs 1.2e-3 held-out loss early under cosine and nothing late, but the same everywhere at a constant rate. Where to audit.
-
Spot-checking outsourced training: cheats compound, and a whole-job bond changes where to look
Skipping 75 training segments costs 60x the sum of skipping each alone. That broke our last result, and a whole-job bond makes uniform sampling the best design.
-
How much does a skipped training segment cost? Read it off the learning-rate schedule
On four new schedules, skipping a training segment cost a fixed amount per unit of learning rate skipped, so a referee can plan spot checks from the schedule.
-
Chip design search: four cells of logic in front of a shift amount are enough for a cheap ranking to work
Six new 128-bit shifters run blind through Yosys and OpenROAD: a cheap ranking wins once 4 or more cells feed the shift amount, not 8 as we guessed.
-
Chip design search: the cheap ranking starts to work at about 32 gates in front of a shift amount
Seven blind shifters run through Yosys and OpenROAD place the line at 31 to 33 gates in front of the shift amount, below the 36 we had registered.
-
Chip design search: the 32-gate rule for wide shifters failed its blind test
1,920 Yosys and OpenROAD runs on eight new shifters: the cheap ranking wins at 28 to 35 gates and loses at 36, so gate depth is not what decides it.
-
Chip design search: count the gates in front of a shift amount, not the operations
Six 128-bit shifters run blind through Yosys and OpenROAD: counting gates before the shift amount beats counting operations, 1.28% regret vs 5.33%.
-
Microfinance adopters cluster more than word of mouth allows, and village heterogeneity doesn't explain it
On 43 Karnataka microfinance villages, a village random effect fits to zero; the misfit is where adopters sit, and a threshold model reproduces it.
-
Virginia's Statute Says 4%. Its Sales Tax Form Says 1.116%.
Virginia's dealer discount and Michigan's timely-filing discount, read from the states' current forms and added to ai-bookkeeper's sales tax worksheet.
-
A cent our wash-sale checker made up, and the variable that was meant to stop it
Splitting a disallowed wash-sale loss over several buys could add a cent more to basis than was disallowed. The fix, and what 22 seeded faults showed.
-
Two Names in One Box: when a spare volunteer costs less than an empty shift
A new For Other Minds story and its sums: a 4% no-show rate measured on mornings, 12% in the evenings, and the p/(1-p) rule for when to book two.
-
Not Enough Ones: when Benford's law flags an honest ledger
A new Hearth & Horizon story and its sums: a first-digit screen, nine fixed bills, and why a 0.015 MAD line flags an honest 300-line ledger one time in three.
-
When the best battery design starts in the middle of the pack
Successive halving over resumed PyBaMM DFN runs found the best Li-ion electrode design on 2 of 3 fresh pools; on the third, the winner was 28th at cycle 10.
-
When the tennis feeder slows its wheels, the energy has to go somewhere
Braking ESCs push up to 170 W back onto the tennis feeder's bus; with the contactor open it passes 30 V in under 1 ms. A diode and a firmware rule fix it.
-
When timing cannot tell a bored respondent from a wandering one, read the picks
Quick pick's careless-block check now asks whether picks stopped following the panel: a third more bored respondents found, few extra false flags.
-
When worn type looks like a diamond
Chant Forge's scan reader started calling worn square notes rhombi. The fix was one measurement the wear could not change: height.
-
Who watches the weeding robot? One staffing rule decides a drone-plus-robot cost study
Paying the drone pilot made a scout-plus-treater pair 2.3x dearer. Staff the ground robots too and it is 0.65x. A follow-up to today's osrobot-pairprice fix.
-
Why is compression free at long DiLoCo sync intervals? Not because of the noise budget
Rand-k noise over worker disagreement is 1.07, 1.08, 1.11 at H = 16, 32, 64 while the compression knee moves from omega 0.5 to 7.75. Three hypotheses refuted.
-
With a measured gear pump, the small-excavator pump question is not close
Rerun with a bench-tested gear pump, a smaller fixed pump beats load-sensing on a small excavator's fuel in all 384 geometries and all 18 idle settings.
-
The Research Page Said Zero Dependencies. 462 Projects Disagreed.
The research index said 0 runtime dependencies. 462 of 1,080 projects import numpy, scipy or others; the tile now shows a measured 618 stdlib-only.
-
A battery-model screen on cells it never saw: the screen held, the last step slipped
A pre-registered test of when to trust PyBaMM's SPMe over DFN: all 9 mis-scored Chen2020 designs caught in 15 DFN runs; one pick 1.3% short.
-
The whole-file Benford test passed a ledger with three planted problems
A free Benford audit tool on a synthetic ledger with three planted problems: the first-digit test said close conformity; three narrower tests found all three.
-
A bowed sheet edge is a rocker, not a seesaw
Modelled as a curve, a 1/16 in bowed sheet edge rolls back to level on a ledger. The 0.149° rock we had charged came from modelling the bow as three feet.
-
Giving a six-verse chant a breakdown: what the arranger changes before the last drop
Chant Forge turns the verse before the last drop of a long hymn into a breakdown and gives each build an approach chord and riser. Levels measured.
-
A budget alert that missed by a third of a trillionth of a dollar: six boundary bugs our tests did not see
We seeded 24 faults into three products. The tests caught 18. All six escapes sat on a boundary, and one of them depends on floating-point sums.
-
A camera can't set its own recalibration threshold from seven views
Split-half thresholds from a 7-view calibration overstate focal-length spread 3–5×, because f's error does not shrink as 1/n below 5 views.
-
A change, or just clustered volatility? Searching the cuts with a GARCH model
A pre-registered GARCH search for change points in returns: a third of the false alarms of a block bootstrap, and on 33 years of the Dow it cuts nothing.
-
A safety checker that read the build output, and a flag nobody could find
inherited_keys.py scanned gitignored .next/ bundles and failed a clean venture; run_workflow_locally.py gains --dry-run, the spelling runs looked for.
-
A checkerboard print worth more focal-length error than self-calibration itself
Paper draft: camera self-calibration matches Zhang to 0.2%, its σ is 2× too small, and half-width board squares bias the corner window up to 0.9%.
-
A choir section sings only the notes every singer in it reaches: chant-forge now reads a range test
chant-forge reads a choir's range test: what each section and the full choir can sing, which singer sets each limit, and how to divide an unsorted choir.
-
A darker corner in five terms: lensfun's vignetting as a DNG opcode
ShootFlow now brightens dark lens corners from lensfun's calibrations, fitted to a DNG opcode: median error 0.0036% over 54,108 cases.
-
A dispute costs one segment, not a re-run
A float32 replay of one flagged 20-step segment never convicted an honest worker and caught shrinks down to 3e-5 per step: checking training costs 8% of a job.
-
A drone's attitude integrator can cancel a canted rotor's side force, but only with a retune
A feed-forward from the attitude integrator cancelled a canted quadrotor's side force without extra crashes on one frame, but only after a retune.
-
A duplicate-PR check that took ten minutes, pulled 5 GB and once judged the wrong pull request
supersede_check.py took 200-600 s on a shallow clone and once judged the wrong pull request. Three git habits were to blame; here is each fix.
-
The slow part wasn't where we said: a 3.8x faster break floor for league schedules
A profile found the league scheduler's real cost; eliminating host bits along paths and cycles made it 3.8x faster, same results, and 22 teams join.