Reversals

The weekly digest for the week of 2026-09-14 was held by the Monday integrity check and released by hand three days later, after one misplaced review verdict was moved to its document’s week, the six affected weeks were re-derived (no status changed, all six negative controls held), and the weekly summary was regenerated so its document total (621, not 664) matched the record.

A GAO report page carries two dates — an internal “Published” date and a later “Publicly Released” date — and the parser took the first, so a late-arriving verdict for that report was filed under a week five weeks before the document’s stored publication date. The integrity check refused to send a digest while a verdict sat in the wrong week. The parser now prefers the release date, and the late-arrival grouping is being changed to follow the stored date so parser drift cannot misfile a verdict again.

Four weekly statuses rose from Stable to Elevated when documents that surfaced weeks after publication were reviewed and counted in the weeks they were actually published: Immigration Enforcement for the week of 2026-09-07 (a DHS inspector general’s unannounced inspection of ICE’s Florida soft-sided facility, in which the department refused all ten recommendations), and Federal Law Enforcement, Independent Agency Rules, and Following Court Orders for the week of 2026-08-10 (Executive Order 14420 on childhood vaccine recommendations and its signing remarks, which direct the Attorney General to litigate against state exemption laws and bypass the federal advisory committee on immunization). The owner reviewed and accepted all four; the Following Court Orders flip is the weakest, since its reasoning concerns federal override of state law rather than the judiciary.

Sources publish late — inspector-general reports appear days after their date and presidential documents reach the compilation of presidential documents about six weeks late. Since 2026-09-14 the weekly ingest re-derives the week a late document belongs to instead of counting it in the week it arrived; this was the first Monday that rule ran live, and these are the statuses it changed.

The July rule that off-topic documents "must not reach … search or exports" is reversed for search (they were never removed from the public dump). Roughly 57,000 stored Federal Register documents that a relevance filter had annotated off-topic become searchable again, fetch-time drops are now stored instead of discarded, and documents no router places are kept under "Not routed to a category". Nothing changes for assessment or statistics: those documents remain outside every count and every weekly status. Explore (the date-sorted browse) shows routed documents by default with a toggle to include the rest; Research (the ranked answer) reads everything.

The corpus and the analysis are independent uses of the record. Topic relevance is an analysis annotation — it says a document is not evidence for a category, not that the document does not exist — while integrity (a body, not a superseded revision) is the only corpus criterion. Hiding annotated documents from search made the public record narrower than the stored one and let a filter mistake become invisible. Measured before the change: no search-latency regression on the dev copy.

The weekly summaries for twenty weeks were regenerated. Their statuses and category lists were correct; the document totals they quoted (for example "volume fell from 660 to 404") no longer matched the record, because the counts underneath had moved. Afterwards the deterministic number check finds no mismatches across the span.

The September relevance filters removed off-topic Federal Register documents from every era, and the same-day whole-term re-derivation moved a few more counts; summaries written before those changes carried the earlier totals. Totals are informational and never drive a status, so the check reports them as drift rather than holding the digest; the owner chose to regenerate rather than leave twenty archived summaries quoting stale numbers.

Five weekly statuses changed after the review verdicts of late-arriving documents were moved to the weeks those documents were actually published: Free and Fair Elections for the weeks of 2026-03-16 and 2026-03-30 rose from Elevated to Confirmed Concern (the executive orders creating the fraud task force and the citizenship-verification requirement now count in their own weeks) while the week of 2026-04-20, which had carried both, fell from Confirmed Concern to Elevated; Government Watchdogs for the week of 2026-08-03 rose from Elevated to Confirmed Concern (the GAO report on FEMA workforce reductions); Spending Money Congress Approved for the week of 2026-08-24 rose from Stable to Elevated (a Treasury inspector general report on IRS compliance trends). The five weeks’ narratives were regenerated to match.

Until 2026-09-14 the weekly ingest stamped every document’s review verdict with the week of the run rather than the week the document was published, so a document that surfaced a few days late counted toward the following week’s status. Four hundred and one verdicts across the current term were moved to their documents’ weeks and every week was re-derived under the repair harness; all six negative controls held. The ingest defect itself was fixed the same day.

The weekly summary was regenerated once and then one sentence of the expert version was edited by hand before the digest went out. The generated text called the week "the largest week-over-week de-escalation in the monitoring period"; the edited sentence says it matches the largest (Christmas week 2025-12-22, twelve elevated categories to two) and, like that week, coincides with a holiday recess. Every number in the summary was checked against the record and stands; the public version carried no such claim and was not edited.

The summary writer sees only the current and previous week, so it cannot check a comparative claim against the term. A regeneration repeated the claim. The deterministic number check verifies counts, not superlatives; a follow-up gives the writer the term series and extends the check.

The weekly digest email was held past its Monday send and released the same day after repair and a one-sentence correction (entry above). The snapshot itself completed cleanly in 72 minutes.

The digest gate found the same derivation-graph violation as the two previous Mondays: two current-term category-weeks carried a document count one off from their score rows because reports the Internet Archive captured late were scored into their true week without that week being re-derived. Both weeks were re-derived under the repair harness (zero status changes, all six negative controls inside threshold). The root cause — every ingest path deriving under the run week — shipped the same day in v1.31.0, together with a finding that the same defect had stamped roughly four hundred review verdicts into the wrong week since February; their restamp is a separate, gated repair.

The weekly digest email was held past its Monday send and released the same day, after repair. The snapshot itself completed cleanly — the first full run since the two memory failures in August.

The digest gate found one derivation-graph violation: three older category-weeks (two in the current term, one in the Biden 2024 baseline) carried a document count one off from their score rows, the late-arrival pattern already tracked as an open bug. Each week was re-derived under the repair harness with zero status changes and all six negative controls inside threshold; the digest was then released to subscribers.

The Biden 2024–25 baseline statistics for all fourteen categories were recomputed from their current weekly aggregates. For twelve categories the values did not move; for Government Information Availability the average weekly document count fell from 168.5 to 11.0, and for Executive Oversight from 67.8 to 62.8.

The relevance filters shipped on 2026-09-02 and 2026-09-04 removed off-topic documents from every era but the derived baseline statistics — which feed the thematic-drift context on category pages, not concern status — were never refreshed. The repair of a baseline week triggered the recompute for this one era under the negative-control gate (all six controls held). The other seven eras carry the same staleness and are scheduled for the same refresh.

A relevance filter removed roughly two thousand routine Federal Register documents from the category’s counts and document lists (Privacy Act notices, paperwork-collection notices, payment and fee schedules, broadcast and spectrum actions, and similar classes that merely mention an oversight office in passing). Zero weekly statuses changed — none of the removed documents had ever driven a detection. Two retrieval queries were also repaired: one had invalid syntax the Federal Register API silently ignored, and one term matched agency citations rather than oversight subjects.

A measurement audit found the category’s Federal Register stream roughly one percent on-topic per document while its confirmed detections came almost entirely from other source types. Because relevant documents here are lexically indistinguishable from routine text, the filter names only the measured noise classes and keeps everything else, erring toward inclusion. Every confirmed document was verified kept, the change was rehearsed with a zero-status-change result before production ran, and the removed documents remain in the corpus under their other categories.

The strongest weekly status now requires at least one confirmed government ACTION: discussion documents (floor speeches, hearings, remarks) still count in full toward "Notable departure" but count at half weight toward "Sustained departure," which additionally demands a primary instrument. Three hundred ninety-five historical weeks whose strongest status rested only on floor speeches were graded down one step to "Notable departure"; no week fell further, and none of the known-event detections were lost.

An audit found more than half of confirmed detections in six categories resting on floor speeches — opposition rhetoric characterizing events rather than instruments of government. The owner chose to anchor the strongest claim to actions while keeping discussions counted. The full-corpus change was rehearsed, the complete week list was reviewed and accepted by the owner before production ran, and every affected week now discloses its evidence mix beside its status.

A relevance filter removed 68,375 off-topic Federal Register documents from the category, 32 misrouted documents were re-assessed under their correct categories, and twenty weekly statuses changed — all downward (nine Confirmed Concern weeks became Elevated, one became Stable, ten Elevated weeks became Stable). Narratives for the affected and active weeks were regenerated.

A measurement audit found the category flooded with unrelated documents (a random sample was on-topic zero times in one hundred) while nearly half its confirmed detections were real signals filed under the wrong category — double-counted there and invisible in their proper place. Every filter decision was validated against an owner-adjudicated sample and a fresh holdout with zero false drops; the exact status changes were accepted by the owner before production ran, and production matched the rehearsal exactly. The removed documents remain in the corpus; only their category assignment changed.

A cross-category audit measured where the system under-finds concerns (audit-sample false-negative rates of 14–25% in five categories) and where confirmations rest heavily on floor speeches (over half in six categories). The findings are tracked publicly and two standing checks were added to the weekly validation so both measures are watched continuously.

Outreach makes detection credibility load-bearing; the audit was commissioned to find both under- and over-finding before readers do. Fixes are scheduled and reported openly rather than applied silently.

The rebuilt salience index (junk case citations and extraction artifacts removed) was accepted with two aspirational pool-hygiene thresholds still unmet: 81 aliases shared across four or more questions (limit 10) and 74 documents recurring in three or more pools (limit 25).

Both conditions predate the change — the previous accepted index measured 76 and 86 — and the metrics this change could have moved both improved (top-ten arm share 20%→18%, mean relevance 0.49→0.50). The cross-question tail the unmet limits measure is separately scheduled work; the owner accepted the gate under its intent and this entry records the choice.

The week’s status changed from consistent-with-norms to notable departure, and its narrative was regenerated, after a newly ingested proclamation was read as a clear departure. The executive-actions narrative for the week of July 13 was regenerated for the same ingest (two monument proclamations); its status did not change.

Sixteen newly approved subject mappings let the Compilation of Presidential Documents route twenty previously skipped documents — emergency-power renewals, monument modifications, deregulatory orders — into the record; one of them moved a published week’s status. The repair ran under the flip gate and full validation.

Fourteen aggregate rows created early for the in-progress week (August 31) were deleted; the weekly snapshot rebuilds them from complete data.

A repair scope mistakenly included the current week, giving it premature status rows — the same artifact corrected on August 28.

The weekly digest email was held past its Monday send and released the same day, after repair.

The snapshot process was killed out of memory before the digest gate ran — and the gate would have held regardless: a late-archived 2024 GAO report had been scored into a baseline week whose aggregate had not been recomputed, one derivation-graph violation. The week was re-derived under the flip gate (zero status changes), validation came back clean, and the digest went out.

The v1.18 retrieval release was accepted with two eval questions one item lower than the baseline (FW2, IM4).

Both lost items were mentions of documents the retrieval had never returned in any run — the reordered pool stopped prompting them. No retrieved document was lost; the owner accepted the gate under its intent and the ledger records the choice.

21 category narratives and 21 weekly summaries were regenerated; 2,030 older narratives were accepted as-is rather than regenerated.

The presidential-document backfill changed the status of 21 category-weeks, leaving narratives that said "no significant anomalies" beside weeks now at a departure status, and summaries whose category counts no longer matched the data.

Twenty-one category-week statuses moved up after the backfilled presidential documents were re-derived: elections (01-12, 02-02, 02-16, 03-09, 03-16, 03-30, 06-01, 06-08), civilService (01-19, 06-01), rulemaking (01-19, 02-09, 06-01), judicialIndependence (02-16, 03-30), lawEnforcement (03-30, 06-01), executiveActions (01-05), fiscal (02-09), military (03-30), mediaFreedom (04-13). All accepted.

Those weeks had held no presidential documents at all while the ingest was stalled; the documents that arrived — Executive Orders 14399 and 14410 among them — were classified as departures.

Presidential documents issued between 2026-01-07 and 2026-08-28 had never been ingested; 1,018 category rows were backfilled, embedded, and reviewed.

The publisher loads these documents about seven weeks after their issue date; the weekly step asked only for the previous week, found nothing, and logged that as success for seven months.

Two hundred current-term verdicts were re-run with the administrations’ names exchanged inside the documents: 11.6% crossed the departure line net of the model’s own noise (19 toward less concerning, 3 toward more). A mirror arm of 190 Biden-era documents renamed to the current administration moved 4.2%.

Verdict rates differ sharply by era; the audit tests whether that is the record or the reviewer. Reading: fragility of borderline verdicts to any renaming, not a party tilt — published on the methodology page.

Thirty-three review annotations (18 current-term, 15 baseline-era) were corrected where the reviewer’s summary misstated what the document said.

A 400-row sampled audit screened the annotations against their documents; 33 of 100 flagged rows were confirmed substantive errors and rewritten from the document text, with the owner’s approval for the baseline rows.

Five hundred ninety-two court-opinion rows that duplicated a revised version of the same opinion were marked superseded (never deleted) and dropped from counts; two weekly statuses moved down as a result (immigrationEnforcement 2026-04-06, lawEnforcement 2026-05-04), and seven moved up when emergency-docket orders were added.

CourtListener issues a second record when a slip opinion is revised; both had been counted. Emergency-docket orders had been excluded by the search default.

Fourteen aggregate rows for the week then in progress were deleted after a repair had created them.

The repair was run through the current date rather than the last completed week, so the overview showed a partial week as the latest.

Two stored weekly summaries were regenerated.

Each named a number of categories and then listed a different number (“Four categories” followed by six). A deterministic number check now runs on every summary before publication.

One baseline-era weekly status changed when GAO reports were added back to 2017.

A new source is required to run uniformly across every period; the owner reviewed and accepted the single flip.

Five review annotations were corrected, beginning with one that said Executive Order 14029 “revokes Schedule F”.

Answer audits found annotations that misread their documents; the corrections were applied from the document text and the sampled audit at scale was scheduled.

The weekly digest was held and released by hand after the missing scores were computed (also on 2026-08-03).

Documents that their sources published late were stored outside the two-week scoring sweep and nothing scored them; the integrity check refused to send a digest built on unscored documents. The snapshot now reconciles unscored documents itself.

Thirty-one baseline-era weekly statuses changed when reports from seven inspectors general were added across every period.

Uniform coverage across periods is required of every new source; the owner reviewed and accepted the flips.

One hundred one baseline-era weekly statuses changed when hearing transcripts were added across every period.

Uniform coverage across periods; the owner reviewed the flips (for example, the 2018 family-separation week) and accepted them as mission-correct.

Sixty-seven documents were dropped from the Press Freedom category by a relevance filter; a fresh holdout week showed no false drops.

Unrelated documents had been routed into the category and were contaminating its counts; the filter was adjudicated document by document (50 of 50 labels agreed) before it shipped.