Evidence record
Data platform architecture & governance
Every published claim on this page was checked by running the command shown. Claims that did not pass were withdrawn before publication; their count is stated below and the full record is available on request.
Measured 2026-08-23 · figures expire with the next publish run; the site is rebuilt from one figures ledger.
Research artifact · published claims
| Claim | How it was checked | Status |
|---|---|---|
| Two independent from-scratch rebuilds of the derived returns panel produced identical output | Two full rebuilds against read-only sources, scratch outside the estate, 465.6 s and 432.0 s; then a row-exact merge comparison of A and B — 17,136,654 identical rows, 0 value mismatches, 0 one-sided rows. | published |
| A staged prediction of the identity fix matched the live production run to the row | Staged rebuild first, then the live full compute (739.3 s). Both runs: null identity 1,664,407 = 9.71%, ambiguous 480, unmapped 1,663,927, beyond-bound-unflagged 0. | published |
| A pipeline cannot write "the company did not disclose X" over a surface with a known hole | Read the coverage gate's refusal path and its docstring: the claim raises GapRefused at write time, there is no override parameter, and a coverage contract older than 36 hours is treated as no contract. | published |
| The estate is 54 databases, 423 tables and 340 indexes across four layers | sqlite_master sweep over all 54 database files, each opened mode=ro, 2026-08-23; layer breakdown by directory — 31 raw, 17 derived, 5 registry, 1 app. | published |
| 397,679,393 rows across 28 named surfaces are under governance | Row counts read mode=ro from each of the 28 row-bearing surfaces and summed, 2026-08-23. The sum mixes grains — filings, facts, positions, speaker turns — so it is a governed-row count, not a fact count. | published |
| Ambiguous ticker identity fell from 2,348 rows across 957 tickers to 28 rows across 14 tickers | The registry assertion that counts ticker rows flagged ambiguous in the identity master, run mode=ro. 2,348 rows over 957 tickers before the fix; 38 over 15 staged; the assertion run at 04:40 UTC on 2026-08-23 recorded 18, and a live re-read after the 06:16 UTC identity build returns 28 over 14. The assertion band is [0, 40]; the figure moves with each build, which is the point of putting it on a timer. | published |
| The defect was reading the wrong validity window, and the obvious repair measured worse than the fix | The master carries two validity-window pairs per identity claim, bookkeeping and evidence; consumers read the bookkeeping pair. The literal fallback proposal was staged first and measured at 16.70% unresolved against a predicted 9.8%, and was rejected on that measurement. | published |
| 30 assertions ran 31 times over 15 distinct days: 776 evaluations, 772 in band, 4 deviations, 0 runner errors | Scan of the assertion run history, 2026-08-09 to 2026-08-23. The registry has a daily timer; the recorded history is 31 runs over 15 days, which is what is claimed here. | published |
| Integrity: 40 databases checked ok on 2026-08-09 — all 40 in scope that day; the estate holds 54 today | quick_check on each of the 40 databases in scope that day, supervised, once. The largest single file, 53.0 GB, took 2,340.8 s at 23 MB/s. There is no recurring readability check across the large files. | published |
12 claims tested · 9 published · 3 withdrawn before publication. Full evidence record available on request — king.hippopotamus@yahoo.com.
Author approval — 2026-08-23, hash b704efae