Stop the deterministic passes from silently truncating the sections they preempt

[sroussey]

Two HIGH-severity correctness bugs on this branch, one commit each. Both are the same shape: a model-free parse replaces the model on a section whose destination the caller has already cleared, so a row the parse never produced is a row the filing loses — with no dead letter, no diagnostic, and the same result on every replay.

1. SKIP_PURPOSE matched its total/source phrases anywhere in the label

parseSpacUseOfProceeds dropped any row whose label contained gross proceeds or offering expenses, which is how most filers qualify the two rows that matter most:

• Underwriting commissions (2.0% of gross proceeds from units offered to public, excluding deferred portion) — the largest expense in the table
• Not held in trust account after offering expenses — the residual trust row

Measured against src/eval/goldenS1Labels.ts: 13 filings, 16 line items lost. The same code path runs under extractor id 424 for the priced 424B1/424B4, i.e. the final deal figures. Example — CIK 1848507, accession 0001193125-21-066104, loses Underwriting commissions (2% of gross proceeds …) = $7,000,000 and Not held in trust account after offering expenses = $525,000.

Every alternative was reviewed, not only the two named:

alternative	verdict
gross proceeds, proceeds after, % public offering, offering expenses	anchored to the start of the label — a total/source/ratio row names itself there
^proceeds from, ^from, ^total, the financial-statement captions	already anchored, unchanged
reimbursed expenses	removed — the golden labels record it as a real line item
per share / per unit	kept as an "anywhere" rule but tested against the label with parentheticals stripped, so Held in trust account ($10.20 per unit) survives while a dilution-table metric row does not

Two follow-on findings the anchoring surfaced:

• One filing (s1_2134856) previously extracted nothing at all — the underwriting-discount row was the only SPAC_USE anchor and it was being dropped, so the whole table failed its own presence check. It now yields 18 rows.
• A filer who factors the sources into a block under a bare Gross proceeds heading writes the children as plain labels (Offering, Private Units). Such a heading now opens a source block that the matching Offering expenses heading closes.

One golden label was wrong rather than the parser: s1_2114227 was missing the Reimbursed expenses row its identically-tabled sibling s1_2114229 carries. Added from the filing (| Reimbursed expenses(4) | | 3,000,000 | | | 3,675,000 |).

2. preempts() authorized a parse on column coverage alone

covers and clears both name destinations. For a multi-row destination a bare table name asserts only "the parse fills every column persist writes" — never "the parse found every ROW". DeterministicPass.complete already expressed the missing claim but was read only for s1:management roster closure, and no wired pass declared it.

Two further live instances beyond finding 1:

• spac-sponsors — covers equals clears exactly. Both prose patterns require our|the immediately before sponsor, so our co-sponsor, Beta Holdings LLC, is … matches neither and an already-cleared spac_sponsor_link was rebuilt with one of two sponsors.
• beneficial_ownership under a full ownershipCoverage — rows failing looksLikeOwner are dropped and stubs carrying a street number are truncated by peelName, before persist and before the section resolves.

complete is now the row half of the contract and a precondition of preempting, checked through a new assertsCompletePopulation (a missing or throwing claim declines, matching how a throwing covers is already treated).

What each pass can honestly claim, and what that costs

pass	claim	preempts?
spac-classification	one row per filing — producing it IS the population	yes, unchanged (14 of 42 fixtures)
sponsor-promote	one row per (extractor, accession); promoteCoverage still answers the column question	yes, unchanged (0 of 42 today)
use-of-proceeds	the walk's own decline log: a labelled row between the first and last line item that carries no readable figure and matches no declared rule	yes, on 16 of the 20 fixtures it parses
spac-sponsors	none — two prose regexes cannot report that the section named no one else	no (was 7 of 42)
beneficial-ownership	none — a walk that filters its own rows cannot tell a dropped row from an absent one	no (was 0 of 42)

Cost/behaviour trade-off. On the committed 42-filing corpus this adds model calls for 7 sponsor sections and 4 use-of-proceeds sections, and removes none elsewhere; beneficial-ownership and sponsor-promote never reached full coverage anyway, so nothing changes for them today. The sponsor and ownership sections are the ones where the honest answer is "this parser cannot assert completeness and must stop preempting" — losing the saving beats losing filed rows.

use-of-proceeds keeps most of its saving because its decline log is a genuine claim derived from the same walk, not a second reading. It errs toward incomplete: three fixtures are declined over sub-table headings that read identically to a blank-figure line item, which costs a model call. The one it declines correctly is s1_2147219, whose Held in trust account row states its amount only as 100.0% of public offering size — golden records it with amount: null, and the parse cannot represent it.

Tests

The recall assertions are the fix's safety net — every pre-existing corpus assertion is precision-only (if (parsed.length === 0) continue plus "does not invent names/purposes outside the golden set"), so a dropped row could never fail CI.

• parseSpacUseOfProceeds.corpus.test.ts — must find every golden line item on a filing it claims to have enumerated (reverting the anchoring makes this red with all 16 lost rows named), plus a floor on how often the completeness claim holds so the predicate cannot degrade into "always false".
• parseSpacSponsors.corpus.test.ts — must find every golden sponsor on a filing it hits; the co-sponsor blind spot is pinned directly, so a stray complete: () => true cannot land as a tidy-up.
• parseSpacClassification.corpus.test.ts — must agree with the golden verdict on every filing it answers. It is the one wired parse whose result is never checked against a model.
• parseOfferingTables.corpus.test.ts — the promote pass must be right about every column promoteCoverage claims, with no null forgiveness (80 field checks across 17 filings).
• parseBeneficialOwnership.corpus.test.ts — the walk's four known drops (AIGH; Acuitas Group Holdings, LLC, Acuitas Capital LLC, Dorado Goose, LLC) are pinned as an exact list. A new gap fails; closing one also fails, which is the prompt to reconsider the claim. Reported rather than fixed here: the pass no longer preempts, so nothing is lost by them today.
• deterministicPass.test.ts / sectionRunner.deterministic.test.ts — full column coverage alone no longer preempts; a false or throwing completeness claim declines.

Two storage tests re-pointed at the new contract (third commit)

Form_S_1.storage.sponsors.test.ts and Form_S_1.storage.ownership.test.ts pinned the superseded preemption contract — the sponsors test asserted the section resolved "without calling the sponsor model", which is exactly what finding 2 removes. Both were verified to pass on base and fail on this branch, i.e. they are this PR's intended behaviour change and not an unrelated break.

They now assert the mirror image: the section reaches the model, the entity is still persisted through that path, and provenance carries the model id and explicitly not DETERMINISTIC_MODEL_ID. Neither was weakened — each now opens by running the deterministic parse over the same section text and requiring it to read the filing outright, so the filing is demonstrably the preemptable-looking case and the point is that it reaches the model anyway. Applied to the base branch, both new files fail there and pass here. No assertion was deleted.

Verification

bun run test           Test Files 428 passed | 3 skipped (431)
                       Tests 3885 passed | 20 skipped (3905), 0 failed
                       (run twice, ~384s each, fully deterministic)
bun run format-check   All matched files use Prettier code style!
bun run build          clean (bun build + tsc)

CI build green on bbb7837.

Correction to an earlier revision of this description. It reported 940 passed, 9 failed and attributed all nine to loads committed S-1 fixtures timing out at 30s under parallel load, calling them pre-existing contention. That was wrong. A full bun run test shows no fixture-loading timeouts at all, and two of those nine were genuine regressions introduced by this branch — the two storage tests described above — masked in the noise of that claim. They are now fixed and the suite is deterministic.

---

🤖 Generated with Claude Code

https://claude.ai/code/session_01Ske1Jwk7fDFxHykfZGEzce

[sroussey]

The two red tests encoded the preemption contract this PR supersedes: both asserted that spac-sponsors / beneficial-ownership stand in for the model on column coverage alone. Confirmed by bisect — both pass on code-extractors and fail at fa62f14.

They now assert the mirror image: the section's prompt reaches the model, and the persisted observation's provenance carries the model id rather than deterministic. Each opens by running the deterministic parse over the same section text and requiring it to read the filing outright, so the filing is demonstrably the preemptable-looking case and the point is that it goes to the model anyway. No assertion was dropped; the replacements fail on code-extractors and pass here. The resale-table test keeps every assertion and gains a name that no longer claims a preempt/no-preempt distinction, since no ownership section preempts now.

Verified on bbb7837: bun run test — 428 files / 3885 tests passed, 3 files / 20 tests skipped, 0 failed (no fixture-loading timeouts; the earlier "9 failed" report was inaccurate and had masked these two). bun run format-check and bun run build clean.

---

Generated by Claude Code