fix(spac): require approval evidence before a general definitive proxy emits a proxy event

[sroussey]

The bug

Form_DEFM14A.storage.ts emitted the proxy lifecycle event for a general definitive statement (DEF 14A / DEF 14C) whenever extractMergerDeal returned a target above the confidence floor:

DEFINITIVE_PROXY_FORMS.has(form) || (GENERAL_DEFINITIVE_PROXY_FORMS.has(form) && extractedDeal)

An extracted deal is not evidence that the meeting approves the combination. mergerDealInstructions() only asks the model to "identify the business-combination target and deal terms", and S1_SECTIONS.BUSINESS_COMBINATION accepts a bare ^the business combination$ heading — so a charter-extension or annual proxy that recites an already-announced deal satisfies the test in full.

The failure chain, end to end

1. A proxy event opens a deal by itself in spacDealGrouping.ts (case "proxy": if (!open) open = openNew(e); open.proxy_date = e.event_date) — no prior Item 1.01 needed.
2. spac8kMilestones.ts then reads pendingMerger = pending != null && pending.proxy_date != null, so the extension meeting's own Item 5.07 maps to vote instead of eight_k.
3. classifyListingRemoval.ts computes approvalDate from a non-null proxy_date alone, so a Form 25 / 25-NSE / Form 15 filed within LISTING_REMOVAL_MAX_DAYS_AFTER_APPROVAL (90 days) classifies completed.
4. Result: a false de-SPAC — terminal status, surviving_name derived and promoted onto current_name, post_merger_* filled — recorded silently for a vehicle that is still searching.

This is precisely the Zalatoris case the proxy_date requirement in CLAUDE.md exists to prevent. The existing negative e2e test used a body with no merger heading at all, so it never exercised this shape.

The fix

A new seeksCombinationApproval.ts supplies deterministic approval evidence, required conjunctively with the extracted deal. Extraction is untouched — the spac_merger_extraction row is still written either way — and the verdict is persisted on it as seeks_combination_approval.

Two line-shaped patterns (line ≤ 300 chars, whitespace-normalized):

• a numbered proposal item naming the filer's defined term — Proposal No. 1 — The Business Combination Proposal, including the SPAC SHAREHOLDER PROPOSAL NO. 1 and THE PENSARE BUSINESS COMBINATION PROPOSAL spellings real filers use — anchored at the start of the line after table/bullet markup;
• a request to approve, adopt, or consider and vote upon the agreement: business combination agreement / agreement and plan of merger / merger agreement.

Deliberately excluded, and the JSDoc says why:

• A bare to approve … business combination — that is the standard extension wording ("to approve an amendment … to extend the date by which the Company must consummate a business combination"). The object of the approval has to be the agreement, not the combination as a concept.
• Any whole-document test for business combination proposal — see the measurement below.

There is no extension/annual exclusion term: a proxy asking for an extension and for approval of the combination is a genuine merger proxy and must still emit.

Why deterministic rather than a model schema field

• Failure costs are asymmetric. A false positive corrupts the primary answer with no trace; a false negative only degrades a fallback, since a real close files an Item 2.01 which maps to completed unconditionally.
• A model field cannot repair existing rows without re-paying the AI bill on every already-processed proxy.
• A required schema field forces an extractor version cycle. Nothing persisted changed, so none is needed.
• The repo already applies "enforce it, don't trust the prompt" for the ownership-subtotal and risk-caption heading guards.

Measurement (STEP 0)

The committed corpus has exactly one proxy fixture (defm14a_sample.txt, hand-authored), which cannot measure anything. I built a real labelled set instead.

Corpus — 348 real EDGAR proxy statements (DEF 14A / DEF 14C) from SIC-6770 filers, discovered through EDGAR full-text search on five deliberately rule-neutral queries ("business combination", "Agreement and Plan of Merger", "business combination agreement", "annual meeting of stockholders", "extension amendment") — vocabulary about agreements and meeting types, never about proposal headings. Every document was rendered through the repo's own parseEdgarHtml + renderMarkdown, i.e. byte-for-byte what the gate reads in production.

Labels — assigned by two signals independent of the rule: the statement annexes the merger / business combination agreement, and it contains unaudited pro forma condensed combined financials (present only when shareholders are voting on the combination itself). That yielded 9 positives / 339 negatives. All 9 positives were then hand-verified by reading their proposal lists (Chenghe, Chenghe I, Aimei Health, MedTech, CC Neuberger II, Investindustrial, TG Venture, Pensare, Cantor Equity Partners III). Negatives were spot-checked across the annex-only band (Churchill VI/VII, Golden Falcon, RF Acquisition, Inflection Point II, FutureTech II, Pure Acquisition) — every one a charter-extension vote.

Result

rule	TP	FN	FP	TN
shipped (2 line-shaped patterns)	9/9	0	0	339

The brief's starting set had to change, and the measurement is why. The proposed whole-text patterns /\bbusiness\s+combination\s+proposal\b/i and /\bmerger\s+proposal\b/i scored 24 false positives out of 339, and every single one was an extension or annual meeting. Two shapes account for all of them:

• the trust boilerplate every extension proxy carries — "a pro rata portion of the funds available in the trust account, as if they had voted against a business combination proposal";
• a cross-reference to a proposal in a different filing — Pure Acquisition Corp's extension proxy says "see the section entitled 'Proposal No. 1—The Business Combination Proposal' located in the Registration Statement".

Making both patterns line-shaped cut that to 2 FPs; anchoring the defined-term pattern at the start of the line took it to 0. That anchoring is principled rather than fitted: a proposal item is a heading or a notice line, so the term begins the line. Both failing shapes are pinned as unit tests.

The line-shaped approve/adopt the agreement pattern from the brief was kept unchanged and carries 8 of the 9 positives on its own with 0 FPs; the defined-term pattern was widened to allow the issuer's own name inside the term (THE **PENSARE** BUSINESS COMBINATION PROPOSAL), which is what recovers the 9th.

Retraction semantics

SpacReportWriter.recordMergerProxy now deletes the proxy event for the accession on the emitProxyEvent === false branch, mirroring the sibling deletes in recordDeregistration / recordUnitSplit / recordCompleted. Reclassification therefore runs in both directions and a replay demotes the deal instead of leaving the old verdict standing. The delete is scoped to that one accession, so it can only retract what a previous run of the same filing wrote — a genuine merger proxy elsewhere in the CIK's stream keeps its event.

Recovery ceremony

mergerProxyDescriptor.filterTodo gains a second, self-extinguishing clause: form in GENERAL_DEFINITIVE_PROXY_FORMS AND an extraction row exists AND seeks_combination_approval is null. Membership is checked before the verdict column, in the spacTrigger8KDescriptor style. NULL therefore means "the gate never ran here" and is load-bearing; a recorded false is a finished answer. The db setup add-missing-column pass creates the column.

sec extractor backfill merger-proxy   # re-derives the verdict; retracts stale proxy events
sec extractor backfill 25-15          # repeat until it reports `processed 0`

Find the affected rows first:

SELECT cik, accession_number, event_date, form FROM spac_event
WHERE event_type = 'proxy' AND form IN ('DEF 14A','DEF 14C') ORDER BY cik, event_date;

SELECT p.cik, c.accession_number AS close_accession, c.form, c.event_date
FROM spac_event p JOIN spac_event c ON c.cik = p.cik AND c.event_type = 'completed'
WHERE p.event_type = 'proxy' AND p.form IN ('DEF 14A','DEF 14C')
  AND c.form IN ('25','25/A','25-NSE','25-NSE/A','15-12B','15-12G','15-15D','20-F','20-F/A');

⚠️ Operators will see status REGRESSIONS on real CIKs as the false closes unwind — completed back to searching / deal_announced, and surviving_name / post_merger_sic / post_merger_tickers / the current_* promotion collapsing back to the spac_* mirror (those five are derived strictly from a completed deal and are never merged forward). This is intended. Every change is captured in spac_history / ChangeLog, and genuinely completed SPACs are re-filled afterwards by sec spac backfill-despac.

No extractor version bump

Persisted extraction rows are unchanged and still correct; only the derived event changes, and it is rebuildable from the document with no model call.

Also corrected in CLAUDE.md

The 8-K item table stated the 5.07 → vote condition as "the pending deal had a definitive-agreement or proxy date". The code requires a proxy date only (pending.proxy_date != null) — which is exactly why a spurious proxy event is enough to fabricate a vote. Now documented as written.

Tests

• seeksCombinationApproval.test.ts (new, 11 cases) — true set covers the numbered heading, the agreement-adoption sentence, a notice of meeting, real TOC rows, and an extension proposal sitting beside a combination proposal. False set covers the real extension wording, an annual-meeting proposal list, a background paragraph reciting a pending combination, the trust-redemption boilerplate, and the cross-reference to another filing.
• Form_DEFM14A.storage.e2e.test.ts — the emitting DEF 14A case now carries an approval-shaped ballot item above the same merger prose (deviation from the brief: the notice line and the segmenter's The Business Combination section heading are two separate lines, as they are in real filings, because the segmenter's heading pattern does not accept … Proposal as a section title); new emits no proxy event for an extension DEF 14A that describes the announced combination (asserts no event, extraction row still written with seeks_combination_approval === false, status stays deal_announced); new a retracted proxy event demotes the deal on replay; new records the gate verdict only for the forms it governs. The DEFM14A/DEFM14C/PRE* cases are untouched by design.
• backfillDescriptors.test.ts — re-selects a general definitive proxy whose gate verdict was never recorded, plus the convergence assertions does not re-select it once the verdict is recorded and does not re-select an M-form proxy, whose verdict is null by design.

Verification

$ bunx vitest run src/sec/forms/proxies-information-statements/ src/task/forms/
 RUN  v4.1.11

 Test Files  21 passed (21)
      Tests  124 passed (124)
   Duration  19.23s

$ bun run format-check
$ prettier --check .
Checking formatting...
All matched files use Prettier code style!

$ bunx tsc --noEmit -p tsconfig.json
exit=0

Full suite: 3674 passed | 5 failed | 20 skipped. All 5 failures are timeouts under parallel load, unrelated to this change: 3 in src/cli/groups/version.test.ts (multi-spawn CLI subprocess tests) which pass in isolation, and parseOfferingTables.corpus.test.ts > loads committed S-1 fixtures, which sits on a timing cliff on this container — 27.6s on a stashed clean tree vs 28.9s with these changes, against a 30s limit — and passes in isolation both with and without the change.

bunx tsc -p tsconfig.test.json --noEmit reports 123 errors, none of them in any file touched here (the suite is documented as red at ~122 and is not in CI).

---

🤖 Generated with Claude Code

https://claude.ai/code/session_01LowBJQsCghLDiHwPN6FgUT

---

Generated by Claude Code