You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
CLAUDE.md is treated as this repo's contract by reviewers and agents, and is unusually good at it. After the 08-17→08-24 window it is measurably behind:
DeterministicPass / preempts / assertsCompletePopulation appear nowhere — grep -c "DeterministicPass\|preempt\|row-completeness" CLAUDE.md = 0. This is the invariant that prevents the data-loss class PR Deterministic passes may not preempt what they cannot supply #311 fixed: covers ⊇ clears for columns, complete for rows, undeclared = decline. The code JSDoc is excellent; the spec is silent. Given that the class shipped this window — tables cleared then refilled with a strict subset, markResolved called, no dead letter, same answer on every replay — this is the single most load-bearing omission.
sec eval management|related-party|spac-sponsors|spac-profile|spac-classification|underwriters|offering-tables|use-of-proceeds|beneficial-ownership — nine commands, ~1,800 LOC: 0 mentions.
daily_index_cursor and the daily-index catch-up leaf, including the "EDGAR 403 means two different things" rule in dailyIndexPublication.ts: 0 mentions.
sec spac process --force <extractor csv> / skip-successful semantics, and sync spacs --only: 0 mentions.
The merger-proxy filterTodo is documented with 2 clauses; backfillDescriptors.ts:233-249 has 3, and the third (the resolved SECTION_NOT_FOUND trace) is what makes it converge.
Two older drifts worth folding into the same pass: the four-cycle-old self-contradicting JSDoc in ExtractorRunRepo, and the two-cycle-old one in classifySpacCandidate.
Found during the 2026-08-24 review. Snapshot: workglow-dev/prd → analysis/grades/2026-08-24/sec-detailed.md.
CLAUDE.md is treated as this repo's contract by reviewers and agents, and is unusually good at it. After the 08-17→08-24 window it is measurably behind:
DeterministicPass/preempts/assertsCompletePopulationappear nowhere —grep -c "DeterministicPass\|preempt\|row-completeness" CLAUDE.md= 0. This is the invariant that prevents the data-loss class PR Deterministic passes may not preempt what they cannot supply #311 fixed:covers ⊇ clearsfor columns,completefor rows, undeclared = decline. The code JSDoc is excellent; the spec is silent. Given that the class shipped this window — tables cleared then refilled with a strict subset,markResolvedcalled, no dead letter, same answer on every replay — this is the single most load-bearing omission.sec eval management|related-party|spac-sponsors|spac-profile|spac-classification|underwriters|offering-tables|use-of-proceeds|beneficial-ownership— nine commands, ~1,800 LOC: 0 mentions.daily_index_cursorand the daily-index catch-up leaf, including the "EDGAR 403 means two different things" rule indailyIndexPublication.ts: 0 mentions.sec spac process --force <extractor csv>/ skip-successful semantics, andsync spacs --only: 0 mentions.filterTodois documented with 2 clauses;backfillDescriptors.ts:233-249has 3, and the third (the resolvedSECTION_NOT_FOUNDtrace) is what makes it converge.Two older drifts worth folding into the same pass: the four-cycle-old self-contradicting JSDoc in
ExtractorRunRepo, and the two-cycle-old one inclassifySpacCandidate.Found during the 2026-08-24 review. Snapshot:
workglow-dev/prd→analysis/grades/2026-08-24/sec-detailed.md.