Composite scoring should respect data_quality_input_corruption

[dackclup]

Background

SPG case from Run #15 (2026-05-10) reveals defense-chain gap:

• data_quality guard correctly suppresses fair_price ensemble when fundamentals corrupted
• But 8-pillar composite still uses corrupted values
• Result: corrupted-data stocks can rank at top of composite list (SPG ranked feat(phase-0): project scaffolding #1 despite obvious data corruption — market_cap $1.62M vs actual ~$76B; tangible BVPS $648,521/share vs actual ~$20-30)

Current behavior

Stage 1: Fetch fundamentals → may return corrupted values
Stage 2: Data quality guard → fires data_quality_input_corruption
         → fair_price methods all skip
         → fair_price.median = null
Stage 3: Composite scoring → STILL uses corrupted pillar inputs
         → produces inflated/erroneous score
Stage 4: Top-5 enforcement → relies on Sloan/Altman/NSI vetoes
         to filter, but data_quality is annotate-not-veto

In SPG's case, the Sloan veto happened to fire (suppressing it from effective Top-5 correctly), but that's coincidental — there's no guarantee data-corrupted stocks always trip another veto.

Proposed fix (Phase 4)

Option	Approach	Trade-off
A	Penalize composite when data_quality_input_corruption fires (e.g., −50% downweight)	Soft signal; doesn't guarantee removal from top
B	Set composite = null (exclude from rankings entirely)	Cleanest UX but loses transparency about the corrupted entity existing
C	Promote data_quality_input_corruption to veto status alongside Sloan/Altman/NSI	Consistent with existing veto framework — corrupted-data stock keeps rank but loses entered_top5 badge

Recommend Option C: most consistent with existing veto framework. data_quality is currently annotate-only because false positives exist; raising bar to veto requires reviewing FP rate first (currently 8/502 ≈ 1.6% — acceptable for veto promotion).

Alternative: Keep annotate, but UI should surface "rank suppressed due to data quality issue" indicator on the corrupted-data stock's rankings card.

Severity

High. Top-of-rankings UX issue, user trust signal. A user landing on / and seeing SPG #1 with $1.62M market cap will lose confidence in the entire ranking layer.

Phase 4 priority

High.

Labels

phase-4, defense-chain, data-quality, ux

References

• Issue Phase-3 fundamentals: shares_outstanding ingested wrong for ~12 tickers #10 (shares_outstanding ingest bug) — likely related root cause; SPG was already in that issue's affected-ticker list (8 tickers fired Step 7.5 sanity guard in Run _avg_3y_roe denominator bug inflates value_trap_risk to 44% of S&P 500 (221 stocks) #11, expanded to include SPG anomaly in Run Fundamentals fetch resilience to SEC API throttling #15)
• Issue _avg_3y_roe denominator bug inflates value_trap_risk to 44% of S&P 500 (221 stocks) #11 (_avg_3y_roe denominator bug) — separate but compounds composite distortion
• PR-3c Step 7.5 sanity guard at compute/valuation/ensemble.py::_has_corrupt_input — currently the only stage that respects corruption signal
• Run Fundamentals fetch resilience to SEC API throttling #15 production output: frontend/public/data/stocks/SPG.json for reference values