-
Pick the shrinkage prior by kind of subgroup, then stop re-choosing it
Choosing each subgroup kind's shrinkage prior from the other survey items agrees with the rule 37 of 39 times, but the two misses erase most of the gain
-
Rotating a survey item off a subgroup's ballot: the paper draft
Seven pre-registered GSS experiments, one argument: keep a calibration subsample, shrink toward the transferred model, borrow only shared drift
-
A shrinkage prior rule, tested on survey items it never saw: half the gain survives
Fixed in advance, the by-kind prior rule cuts subgroup error to 0.85 of the zero prior's on 7 fresh GSS items, against 0.65 on the items it came from
-
The citation checker now reads blog posts without leaving a stray file
cite_check.py wrote an untracked cache beside any post it checked and crashed on a directory. Posts now borrow the 21 project caches instead.
-
Letting a simulated arm rest against a soft side: 28.5 degrees down to 18
The drape's rigid body stopped the arms at 28.5 degrees. A side carved round the arm brings them to 18, and all nine test drapes now settle.
-
Complex centrality, the seeding measure for complex contagions, does not beat degree discount on 43 villages
On 43 village networks, the best complex-centrality rule loses up to 3.7% of adopters, degree discount 1.4%; seeding whole neighbourhoods loses 10 to 13%.
-
Complex centrality helps more as the fitted adoption threshold rises, and still loses to degree discount
Over 500 refitted bootstrap draws on 43 village networks, complex centrality loses a median 3.3% of adopters; degree discount 1.0%, and it wins every draw.
-
A floor that fires only on thin pilots: it fixes the cells it can reach, and good coders pay for it
Pre-registered, 114 cells: flooring only pilots with 3 or 4 coder misses repairs 4 failing cells but keeps 79% of the gain; good coders pay most.
-
Curitiba's ISS rates: the city where 2% is always a legal rate
Curitiba's ISS law sets 2%, 2.5%, 4% or 5% by service. We added it to the NFS-e checker and tested it on 19 public files. Two clauses limit what it can catch.
-
1,436 imports our cycle checker couldn't see
cyclehound now follows tsconfig path aliases like @/lib/x, and TypeScript 5.9's own resolver sends the same 1,436 imports to the same files.
-
Dating each fitted leak source by its first day did not find the slow leaks
On BattLeDIM's L-Town, a daily tracker that dates fitted leak sources localises 8 of 32 leaks, below a one-week rise rule on the same data (14).
-
Deciding each test as it arrives: online false discovery rate control, measured
LORD++, SAFFRON and LOND now decide streams of tests one at a time. A seeded simulation shows when that costs power against batch BH, and when it does not.
-
DiLoCo's clean-last-sync gain is noise, not sparsity, and it is biggest at the lightest compression
Rand-k at keep 1/4 on every sync but the last beats uncompressed DiLoCo by 0.018 nats; Gaussian noise does the same as rand-k; keep 1/16 stops paying.
-
We never measured our drone's motor lag. Sweeping it keeps the rankings but moves every crash count
A 4x sweep of an unsourced motor-lag law keeps our quadrotor co-design rankings but changes crash counts and which designs pass a stability check.
-
Eighteen nanoseconds a segment: where a changepoint test spends 14 seconds
One changepoint test scores 135 million Poisson segments. Inlining the cost cut it from 16.1 s to 14.0 s with identical answers; the rest is the algorithm.
-
Eleven Degrees to the Glass: a story about frost that sits on one side of a solar roof
A new Hearth & Horizon story: both slopes of a chapel's solar roof frost on clear nights. The south clears by 9:15, the north after noon.
-
Eleven WebGL frames for three pictures: the shed designer's thumbnails drew four times each
The shed designer drew 11 WebGL frames for 3 pictures. Off-page canvases now draw once, at the snapshot: same pixels, a quarter less GPU CPU in tests.
-
Exact cash and a bottle of water: testing a receipt fix on the photos it was written for
A rule that stopped AI Bookkeeper's receipt reader taking the cash handed over as the total fixed one photo and broke another. Here is the second fix.
-
Fair workweek pay lives in the change log, not the schedule
Shift Roster Builder's self-hosted hub now logs every change to a posted week and prices predictability pay under Seattle, Oregon or Chicago rules.
-
We cut 17 seconds from two test files and the suite got 2 seconds faster
Two league-scheduler test files lost 17 s of repeated searching. The whole suite moved 2 s, within noise, because it was already close to CPU-bound.
-
Five hundred keyword rules, asked the same question on every import
ai-bookkeeper tested every bank line against 499 keyword regexes on every import. Remembering the matches halved its slowest test and gave identical answers.
-
Five tool pages opened with a 120-to-180-word paragraph. Now none can.
The paragraph under a tool's headline is what a visitor reads first. Five ran past 120 words. Cut to 44 to 106, and a test now stops any page going over 120.
-
Flip every host: a free factor of two in the break floor, and why 30 teams still wait
Swapping home and away for every game leaves a season's breaks unchanged. The schedule generator now uses that to halve its tables. 30 teams are still too slow.
-
For Other Minds: Laetare with Laetare, a story about the Easter that moved
A parish's plate collection looks up 52% in March and down 31% in April. Easter moved. A short story about comparing Sunday with Sunday.
-
A 40-team league in three sessions a night, solved by writing down one night
A SAT model that timed out at 300 s for 40 and 46 teams solves in 0.1 s once a cyclic renaming turns one night into the whole season.
-
Fourteen thousand pictures nobody looked at
Our site-builder tests took 403 s on a fresh machine, mostly drawing share cards for fixture builds no test reads. Two small changes cut it to 230 s.
-
Why DiLoCo compression looks free at long sync intervals: a training gain cancels the last round's cost
Forking 16 runs at their last sync shows rand-k noise helps training at H = 64 (-0.014 nats) while the final round costs +0.012. The two cancel.
-
Ten of eleven flagship rows sit within 80 characters of the cap, so the edit tool now makes room itself
Every flagship next: field sits near its 2,000-character cap and edits took two or three tries; flagship_field.py --make-room lands them in one
-
Making a learned demosaic twice as fast without changing a single output bit
We fused a learned Bayer demosaic's 30 feature passes into one row pipeline. It runs 1.8-2.2x faster and matches the numpy reference bit for bit on seven raws.
-
Reading a gear pump's drag where the excavator optimum picks it clears the 20% bar, at a price
Averaging gear-pump idle drag over the sizes the fuel optimum picks cuts the rule's out-of-sample error 29%, but it does worse on catalogue pumps alone.
-
A 4x faster .gitignore check by copying a shortcut git already takes
The .gitignore Explainer took 2.7 s on 200 rules and 20,000 paths. Git skips its matcher for rules like *.log; doing the same got it to 0.7 s, output unchanged.
-
In a .gitignore, ? matches a byte, not a letter: café.txt slips past caf?.txt
Git matches .gitignore wildcards byte by byte. é is two bytes in UTF-8, so ? and [...] see half a letter. Our explainer got this wrong until we tested it.
-
Checking one step of a Gödel-style proof of God is NP-complete when properties have complements
Without rigid positivity, the frames where Anderson's essence step always holds are polynomial to test, or coNP-complete once every property has a complement.
-
A guard fixed one camera our boosted demosaic broke, and missed two others
A fallback to the linear model saved a Canon R8 and passed six fresh cameras, but two held-out cameras slipped past it, so the trees stay unshipped.
-
A half-cent loss printed as a gain, and 30 formatters that printed -$0.00
Mutation testing a delivery-fee analyzer found a money formatter that showed a half-cent loss as a gain, and 30 more that printed -$0.00.
-
Survey half a village in two visits and you can seed a rollout almost as well as with a full network census
On 43 village networks, a two-stage survey of half the households plus degree discount loses 1.6% of adopters; a full census 1.4%, a random half 3.3%.
-
Half the tiles our planner checked were not on the floor
Our tile planner checked 43 million off-floor tiles in one test run. Skipping only those it can prove miss made it 31% faster, with byte-identical output.
-
The bad cloth states were a rim-fill bug: moving a drape between meshes without breaking its edge
Copying a cloth drape onto a coarser mesh left a kink at its edge. Extending the edge along the cloth fixed 18 of 19 bad transfers.
-
HDBSCAN in the browser, and the 96 labels that depend on how you sort a tie
A new free HDBSCAN tool matches scikit-learn on 5,712 labels once one sort is made stable. Without that, tie order alone moved up to 96 of 300 labels.
-
Hear when your Stripe card declines jump, from your own GitHub account
A free watch that reads payments, revenue and declines from your own Stripe key and messages you when one moves to a new level. Synthetic shops: 20 of 20 heard.
-
Hearth & Horizon: Six Times for a Second
A new story about a clinic generator sized from nameplate watts, a borehole pump that pulls six times its running current to start, and the fix.
-
Hearth & Horizon: The Gallery That Is Full for the First Time
A new science fiction story about a church gallery rated per square metre, a head count of 146, and why everyone at the rail bends the joists hardest.
-
Hearth & Horizon: What the Night Air Lends
A new story about drying maize with air. Its model overturned the planned fix: switching the fans off on foggy nights made the top of the bin dry later.
-
How many seconds of video does a robot camera need to self-calibrate its focal length to 2%?
Pooling 1.2 s windows of TUM RGB-D video: a random-effects pool keeps the error bar honest, independence overclaims 2.8x, and 2% takes about eight windows
-
How stiff must a nailing robot's joints be? About 10,000 N m per radian, if the nose grips the sheet
Joint windup under a nailer's 25-50 N press costs a G1 5.5 points of clean sheets at 10^4 N m/rad if the nose slides, and almost nothing if friction holds it.
-
The Front Page Now Shows the Tools, Not Just Stars
At desktop width a third of the hub's hero was empty starfield. It now shows three real tools with their worked examples loaded, each one a link.
-
One calibration is not enough: a humanoid nailer's backlash bias moves 3.4 mm across the wall
Held joints' backlash is a fixed bias only at one pose. Across a G1's reach three joints flip their load, leaving nails up to 3.4 mm off after one calibration.
-
Compensating a humanoid nailer's backlash works only if the compensator knows the balance controller
Predicting each held joint's side removes a G1's backlash offset only with the robot's own balance controller in the model. A wrong one leaves 1.2 to 1.9 mm.
-
A humanoid pressing a nailer into a wall has to lean into it, and that undoes last milestone's flips
Through real soles, a Unitree G1 standing upright balances only 26–39 N of press above 0.8 m. Leaning fixes it, but backlash still needs per-pose care.
-
Gear backlash in a humanoid nailing robot: the joints that matter are the ankles
Backlash only adds error on joints nothing loads one way. Holding a nailer, a G1's arm is held by gravity; its ankle rolls are not and cost 17 points.