docs(phase4.5d): mark 4.5d wave done + tighten CLAUDE.md phase block

[dackclup]

Summary

Triple-doc lockstep bump after PR #97 (Phase 4.5d — earnings-quality time-series) merged + production-verified on workflow run #50.

• PHASE_STATUS.md — row 4.5 → "4.5a + 4.5b + 4.5c + 4.5d waves complete"; defense layer 9 → 16. 4.5d ✅ DONE block added with production fire rates + design-substitution note. Phase 4.5 timeline summary table 4.5a-d rows flipped ⚪ → ✅.
• SKILL.md — Rule 16 active-veto count 5 → 7 (added beneish_manipulation_veto + dechow_manipulation_veto). Schema versions table appended with a 4.5a-d wave row noting zero data-shape delta (all 9 new flag identifiers are list[str] additions on existing fields).
• WORKFLOW.md — 4.5d task list [ ] → [x] with PR feat(scoring): earnings-quality time-series (PR 4.5d — accruals_momentum_high + loss_avoidance_pattern) #97 + production verification snapshot + design substitution (m_score_deteriorating → accruals_momentum_high).
• CLAUDE.md — Phase status section trimmed ~65 → ~12 lines (now points at PHASE_STATUS.md as canonical). Layout-table veto count 5 → 7, test count 526 → 831. Closed issue Investigate NVDA Sloan accruals flag — genuine red flag or growth-stage artifact? #7 removed from Gotchas; loss_avoidance_pattern universe mismatch added.

Production-verified snapshot

Field	Value
Commit	c3b29af4 (chore: update rankings 2026-05-17)
Workflow run	#50 (id 25982432928)
Wall clock	6m24s (warm cache)
Schema	0.7.1-phase4g (unchanged — 4.5 wave is annotate strings only)
Universe	502
Active vetoes	7 (sloan_accruals_top_decile · altman_distress · net_issuance · beneish_manipulation_veto · data_quality_input_corruption · dechow_manipulation_veto · non_reliance_filing)
Defense layer total	16 (was 14 after 4.5c, +2 from 4.5d annotates)
Tests	831 offline + 17 @network

4.5d production fire rates

Flag	Fired	Predicted	Status
accruals_momentum_high	50 / 502 (10.0%)	3-8%	⚠️ slightly hot, within annotate-only band
loss_avoidance_pattern	0 / 502 (0.0%)	1-3%	⚠️ S&P 500 universe mismatch (Burgstahler-Dichev 1997 cohort thresholds NI ≤ $5M / EPS ≤ $0.05 too tight for large-caps)

The loss_avoidance_pattern zero-fire is a calibration / universe-scale issue, not a bug. Annotate-only so the zero-fire is safe-by-default. Phase 4.5 follow-up draft staged at /tmp/issue_drafts/issue_loss_avoidance_universe_mismatch.md proposing universe-aware thresholds + AAER-cohort calibration validation.

Test plan

• ruff check . — clean
• All 4 docs edited in single commit (lockstep invariant)
• PHASE_STATUS.md production-verified commit SHA matches metadata.json::git_commit (c3b29af4)
• SKILL.md active-veto count matches defense-scorecard / production risk_flags inventory (7)
• CLAUDE.md test count matches pytest --collect-only -q output (831)
• No code under compute/, frontend/, tests/ changed — pure docs PR

Next decision

After this merges, pick:

1. 4.5e — Form 4 insider clustering (insider_sell_cluster + c_suite_unusual_sell, ~420 LOC, new SEC parser, ~3 weeks)
2. 4.5f — manipulation_index 0-100 composite + composite penalty + UI + schema bump → tag v1.2.0-phase4.5 (~250 LOC, ~1 week)

Recommendation: jump to 4.5f to ship a tagged v1.2.0 milestone first; revisit 4.5e once Form 4 ingest is needed for a separate use case.

https://claude.ai/code/session_015649aRyi2bvciQYZVNACd2

---

Generated by Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
quantrank	Ready	Preview, Comment	May 17, 2026 5:54am