Every error this project has caught, dated, with what would have been published had the gates not held, and what changed so the class cannot recur. An institution that hides its errors gets caught; one that accounts for them gets cited. Newest first. Entries are append-only.
The first draft of source_tiers.json designated four broadcast
outlets tier 4 with specific fabrication claims the project could not
cite. Caught at the gate before commit. Published version holds those
outlets at tier 3 pending citations; tier-4 designations now require
evidence the project itself can produce or cite. No public exposure.
The first run of the episode-actors decomposition ranked any busy India dyad as an episode "driver": a Pakistan-channel episode showed Ukraine as its top driver. Caught by the too-weird sanity rule before push. Fixed with registered channel-partner lists; unrelated active dyads now appear only as labeled context. No public exposure.
The comparator publisher's first output labeled a half-finished week with a complete week's date; India's "latest" read 10.8 where the true complete-week value was 42.4. Caught in pre-push verification. The publisher now drops any trailing week whose label postdates the data. No public exposure.
Not an error, recorded here as precedent: the stress gauge's pre-registered detection rule scored 1 of 21 episodes and was published unchanged. Registrations bind; low numbers are findings.
The first geometry bake produced zero countries: Douglas-Peucker on closed rings degenerates (start equals end) and collapsed every polygon to two points. Caught in browser verification before commit. Fixed with a farthest-point anchor. No public exposure.
Pre-existing defect found during verification: the analysis page's chart code read CSS variables its stylesheet never defined; the attention-pricing-gap chart had never rendered for any visitor and every other analysis chart drew in fallback grey. Fixed the same hour with the palette added and cache-busters bumped. Public exposure: cosmetic-to-broken charts, no wrong numbers.
The daily pipeline computed 2026-07-30 correctly and lost the final push to a race with the backfill chain; a conflicted rebase swallowed by an error guard discarded the day while every step reported green. Fixed with fail-loud push retries and single-writer file ownership. Public exposure: the site showed the previous day's data, correctly labeled, for several hours. No wrong numbers.
"Galwan River" sat in the china_east Wikipedia list while the Galwan clash is a validation episode, which would have made that detection circular. Caught in review; the ex-ante CI test now covers every query list in the repository. No public exposure.
Before this ledger existed, a misaligned market window once produced 53 statistically significant event-study results from real data. The too-good-to-be-true sanity gate caught it before publication. That incident is why every result that flatters the project is treated as a bug until independently recomputed.
The daily workflow updated the events store in its workspace and published map aggregates from it, while a stale exclusion (left over from the finished backfill chain's race protection) kept the updated store files out of the commit. Published partner totals exceeded the committed store by 29 to 65 events for about a day; derived shares were correct for the data used, but the published numbers were not reproducible from the committed raw files. Caught by the first run of the dual-computation audit module, the same day it was written. Fixed: the exclusion is removed, and the audit now runs inside the daily workflow before any commit. Public exposure: map totals ~0.03% high relative to the reproducible store, one day.
The dual-computation audit's gauge check compared two independently rounded published values under a 0.051 tolerance, but rounded-inputs arithmetic can honestly differ by up to about 0.10, so the audit false-positived (54.9 vs 54.85) and failed three daily runs across 2026-08-02 and 08-03. By design the failure was loud and the site served stale-but-labeled data, never wrong data; the availability cost was real. Tolerance corrected to 0.101 for that one check with the derivation in a comment. Also armored the same day: the daily schedule's crons have never fired on time in this repository, so the reliably-firing nowcast workflow now dispatches the daily run whenever yesterday's final is unpublished past 00:30 UTC.