-
RapidOCR now runs in the browser: the receipts page reads 184 of 196 photo totals, up from 155
A JavaScript port of RapidOCR on ONNX Runtime Web brings PaddleOCR to ai-bookkeeper's page: photo totals 155 to 184 of 196, taxes 135 to 166, nothing uploaded.
-
RapidOCR reads the receipt totals. Tesseract, read four times after it, still finds the taxes
Voting RapidOCR's reading with tesseract's four kept totals level and added taxes: 28 to 32 of 41 on held-out photos, 142 to 162 of 185 on US receipts.
-
Moving in-browser OCR to a Web Worker: the receipts tab's freeze goes from 3.3 s to 51 ms
ai-bookkeeper's receipt-photo OCR now runs RapidOCR in a Web Worker beside tesseract.js: the same text on 200 photos, 30% faster, and the tab no longer freezes.
-
Reading Deputy rosters when Deputy's API docs won't load
Shift Roster Builder now reads Deputy rosters straight from its API. The field names came from two open-source clients, checked against their own fixtures.
-
The receipt total that was right until the checker fixed it
A Denny's slip read perfectly, then corrected from 5.29 to 4.29 because the gratuity line was in no check. The in-browser receipt reader now gets 189 of 196.
-
When the receipt's amounts sit one row below their words
RapidOCR boxes a receipt's words and amounts apart, and rows slip by one. A subtotal-plus-tax check puts them back: US photo totals 182 to 187 of 196.
-
The receipts page dropped every rupiah tax its five readings agreed on; now it keeps 205 of 267
Writing a voted tax back as "Tax 8728.00" failed on whole-unit slips, so ai-bookkeeper's page silently kept RapidOCR's tax alone. Fixed: 177 to 205 of 267.
-
Our receipt-photo rules were tuned to 100 photos. On 200 new ones they read 59%, not 70%
Every receipt-photo rule was chosen on 100 CORD validation photos. Unseen train photos read 59% of totals, not 70%. New rules: train 62%, test 58%.
-
Our receipt reader on 472 receipts it was never tuned on: 88% of totals, and why taxes lagged
On WildReceipt's held-out photos the free, offline receipt reader got 88% of totals right. Glued tax words like MwSt19% hid taxes; fixing them added 10.
-
Five OCR readings agreed the tax was 1.86, and the checker changed it to 1.56
With its own RapidOCR port leading, the in-browser receipt reader gets 190 of 196 US receipt totals. One lost tax was a fix that ignored a gratuity.
-
Where our receipt reader still lost the tax: Target's state lines, VAT tables and 'total includes GST'
633 fresh receipt photos showed where tax went unread: Target state lines, VAT tables, GST-included lines. Big gain on them, a small one held out.
-
A receipt that says 6.008 almost always means 6.000: 600 more photos, two fixes, one dead end
600 unseen CORD receipt photos read 61% of totals. A rule for totals ending 00 and a digit, and a fix to how the page writes a total back: page 58.5% to 61.2%.
-
Regression discontinuity in a browser tab, and what the default bandwidth costs in coverage
A free RD tool matches rdrobust within 1e-8 at fixed bandwidths. On 500 shared simulated samples its own bandwidth covered 93.2% against rdrobust's 91.8%.
-
Revit writes Reference as a label, and an IDS check asking for an identifier fails every object
On two real Revit exports, 791 of 966 objects failed IDS checks only because of a data type. The free IDS editor now offers to drop the dataType.
-
Romania, Portugal, Croatia and the Netherlands: our e-invoice rules against Saxon
Four national e-invoice rule sets, run in the browser and on Saxon over 5,556 test copies: no rule ids differed. The one mismatch was in our test script.
-
Two seconds past the budget: what a scheduler's last search stage buys
Shift Roster Builder's day sweeps run 2 s past a spent 20 s budget. Logged gain by gain on 23 builds, the last 1.5 s covered 2 of 11 open hours. They stay.
-
Six of France's 43 invoice-status rules cannot fire
A French invoice status with no invoice number passes all 43 BR-FR-CDV rules on Saxon and on our engine. Why six rules are dead, and what we check instead.
-
Who gets matched with whom? A Gale-Shapley tool, and the search that visited some answers twice
A browser tool runs Gale-Shapley with capacities, lists every stable matching and checks one made by hand. Its listing had a bug that showed only at scale.
-
Staggered difference-in-differences in a browser tab, and the year before adoption
A free DiD tool matches R's did to 1e-9. On castle-doctrine data, TWFE says 0.069 and the clean estimator 0.019; the year before adoption explains most of it.
-
Stopping a refused trial early: why a lower bound did not speed up 30-team league schedules
A search for 30-team round robins refuses most moves by one to five breaks, so bounds that stop a trial early saved 6% of steps and no time.
-
Proposition 99's synthetic California is matched on sales, not on income
Covariate balance of the outcome-only synthetic control: California's synthetic twin is 1.66 sd poorer, and Massachusetts is balanced yet still not found.
-
Which covariate imbalance biases a synthetic control? Ask the placebos, not the 0.25 rule
A pre-registered simulation: the 0.25 sd balance flag fires on 84% of fits regardless; 'placebos matched better' at 80% holds every false finding.
-
Our e-invoice CLI skipped a check the web page ran: 24 of 34 files read differently
The Python CLI read party ids only from JSON, so Peppol's id format rule never fired on XML. A two-reader diff found 24 of 34 files apart; now 0.
-
Asking for dates a building model already holds: a date range from every value
The IDS editor now offers the range a model's dates span. With 600 doors, the earliest date was not among the 50 values on screen.
-
The greenhouse planner was leaving 200 peppers off a bench they fit on
Greenhouse Bench Planner grew no peppers on its own example. Trying other booking orders lifts its margin from $3,802 to $4,548.
-
The incremental cost function that made our scheduler twice as slow
Recounting a 12-team league's carry-over value in full on every annealing move beat updating it incrementally; the win was in the venue search instead.
-
The slowest recurrence rule is the one that never happens
456 of 3,200 test expansions in our RRULE explainer found no dates and took 92% of the time. Checking only possible days made them 5x faster.
-
A tennis ball is in the wheels for 6 ms, and the feeder's drop reading noticed
Spreading a ball's take over its contact time raised the drop readings' scatter up to sixfold. A contact-time term fixes it; the fit can pick the time itself.
-
The tennis feeder's reading floor was a search grid, not the physics
A ten-times-finer simulator left the 0.3% drop-reading scatter unchanged. The fit's grid for the ball's moment was the floor; one extra solve halves it.
-
The tennis feeder's spin error was there without the jitter
Plants with no tach jitter miss the spin fit by the same 0.5%. Timing every edge exactly barely moved the readings, so the 1 us timer is not the floor.
-
The tennis feeder's unsure marks find the noisy wheels, and weighting them down still doesn't help
The feeder firmware marks wheel readings it is unsure of. On 72 simulated builds the marks found the noisy wheels, but down-weighting them gained nothing.
-
The tennis feeder's wait for slow wheels made one shot in three late
A held-out test kept the feeder's 1,000 rpm fit floor and found its wait for slow wheels made shots up to 1.7 s late. An opt-in fix removed the delay.
-
The tennis feeder's worse result on a jittery tach was one seed's draw
One held-out seed made a firmware setting look worse on a jittery tach. Six seeds, 3,888 wheels and a check of each wheel's fit showed it was chance.
-
Thirty amps at 12 volts needs 4 AWG over 15 feet: a free voltage drop calculator
A free voltage drop calculator: every wire size at once, the smallest within your target, the longest run and the watts lost, with computed resistance.
-
Three OCR readings agreeing on a receipt total is not a check
Tesseract's four readings outvoted RapidOCR's right receipt total 19 times on 583 photos. Requiring the receipt's own arithmetic took totals from 526 to 542.
-
Twelve security checks, one wait: workflow-audit's tests in a third of the time
workflow-audit ran twelve 30 ms checkers one after another. Running them at once cut its tests from 8.3 s to 2.5 s with byte-identical output.
-
Two elements, one GlobalId: what 163 public IFC files hold
IFC says no two objects share a GlobalId. 3 of 163 public models break that rule, and finding them exposed a parser bug in our IFC reader.
-
Re-scoring four receipt tax rules, and the VAT table they still read wrong
Four receipt tax rules shipped unscored. Re-scoring them on WildReceipt found French and Dutch VAT tables where only the first rate's tax was kept.
-
When a schedule search stalls: 47 teams, five runs, no climb past 13 of 16 nights
Five resumed searches spent 26,870 core-seconds on a 47-team, 8-court season and none got past 13 of 16 wait-free nights. What that tells the next attempt.
-
Alpha's confidence interval gets worse with more respondents, when items are unequal
A new free reliability tool, and a simulation: with unequal items, alpha's 95% interval covered the true reliability 11% of the time at n = 1,000.
-
Who pays for the shared road? A Shapley value and nucleolus tool, checked against the Talmud
A browser tool splits a shared cost by the Shapley value and nucleolus, checks each split against the core, and reproduces the Talmud's estate rulings.
-
29 and 35 team leagues with no waits: re-dealing two nights at a time
29 and 35 team leagues, three games a night, used to leave hundreds of waits. A search that re-deals two nights at a time with SAT got both to zero.
-
A 41-team league with no waits: the last night was impossible, so we changed who plays on it
41 teams, 7 courts, three back-to-back games a night. One night's pairings could never be laid out. Re-dealing it with the short last night fixed the season.
-
A 41-team league, 13 nights of 14 wait-free, and a search that can pick up where it stopped
A neighbourhood search got 41 teams on 7 courts to 13 of 14 wait-free nights. Two fixes made that possible: the search keeps its own layouts and can resume.
-
A ball machine's 1% speed bias was 0.7 grams of tennis ball
Our open tennis ball machine read its own launch speed 1-2% high on every simulated build. The cause was two ball masses in one codebase.
-
The bookkeeper knew hotels but not trains
We added 1,193 public transport networks from OpenStreetMap to an open-source categorizer. It fixed one held-out line and broke none. Here is why so few.
-
Checking IFC models against IDS in a browser tab: 334 of 334 buildingSMART test cases
A free tool checks IFC models against buildingSMART IDS files in the browser. It gets all 334 official test cases right and matches IfcTester on real models.
-
A card copy that only took the raws it could develop left 6 of 71 behind
ShootFlow's card copy skipped raw formats it could not develop. It now takes 27, lists what it leaves, and develops Leica RWL files.
-
A chant-reader fix that worked on its training set and failed two held-out tests
A chant consonant reader missed nasals on high short notes. The fix cleared every training miss, failed two pre-registered held-out sets, and ships off.
-
A chant-scan fix that was right, safe, and never fired on the held-out pages
Chant Forge's scan reader misread a liquescent sitting on a staff line. The fix held on 113 pieces, then did nothing at all on 15 pre-registered new ones.