Summary
Phase 19 ran five hardening tracks against the locked Phase-18 result. Four tracks confirmed or refined the finding; one — the baseline comparison — produced an uncomfortable result: two single primitives slightly outperform the A/G composite on AUROC. The honest Phase-19 summary preserves the corporate signal’s legitimacy but downgrades the strongest version of the composite-vs-primitives claim.
The five hardening tracks
1. Anchor refresh (out-of-time)
A bounded EDGAR FTS refresh from 2026-01-01 → today found 6 new high-confidence Item-1.03 anchors; 5 of 6 are already in the Phase-18 panel. The 22-row out-of-time descriptive check is consistent with Phase 18 in direction (mean percentile 56–62 for structural distress) but is too small to support a statistical claim. Full writeup: Anchor Refresh Note.
2. Sector stratification
Pooled AUROC 0.711 is broadly distributed across the reliable SIC divisions (Manufacturing-2 0.729, Wholesale/Retail 0.742, Manufacturing-3 0.639). The Manufacturing-3 (durables / electronics) cell is meaningfully weaker than the pooled headline. Smaller cells are flagged DESCRIPTIVE_ONLY or NO_STRONG_CLAIM. Full writeup: Sector Analysis.
3. Calibration / monitoring bands
Predeclared (not tuned) quantile bands on the structural-distress test-fold distribution. The Watch band (80–95th percentile) is the most informative tier (event rate 3.1%, lift 2.6×, recall 39%). The top-1% Critical band shows a tail dropoff (event rate 0.4%, lift 0.37×) and is noise-prone. Full writeup: Calibration Policy.
4. Baseline comparison (uncomfortable)
Two single-primitive baselines slightly outperform the A/G composite on AUROC: operating cash flow / assets 0.736, interest coverage 0.726, vs structural distress 0.711. The composite’s value is robustness and regime separation, not maximum raw discrimination.
Full writeup: Baseline Comparison.
5. Divergence-anchor study
Within the divergence-state cohort (firm-years where A > 0 and G < 0), the canonical R / divergence geometry posts AUROC 0.705 — beating structural distress (0.666) and joint-collapse severity (0.468 chance) for future structural deterioration. This is first internal evidence that canonical R behaves as predicted for its intended divergence / masking regime. The anchor is internal future-state, not external. Full writeup: Divergence-Anchor Study.
Two-label decision
- Process label
- PARTIAL: Phase 19 completed but baseline/sector results weakened the claim
- Scientific verdict
- corporate signal survives but is weaker than simple baselines
Why "PARTIAL" and not "ADVANCED"
- All five hardening tracks shipped end-to-end (no blockers).
- But the baseline comparison returned the uncomfortable finding that single primitives match or slightly beat the composite on AUROC.
- Sector stratification revealed Manufacturing-3 as a noticeably weaker cell.
- Combined, these weaken the "the composite adds architecture" claim without invalidating "the signal is real."
What stays locked
- All Phase 1 / 7 / 8 / 11 / 12 / 13 / 15 / 16 / 17 / 18 / 18A specs and outputs (20 baseline artifacts byte-identical).
- Phase-13 geometry definitions; Phase-11 winsorize + robust-z; coverage gate at 0.75; λ ∈ {0.0, 0.5, 1.0} with corporate primary at λ = 0.
- The fetch-completion gate added in the Phase-18 operational correction.
Recommended next focus
Higher-ed Task #15 (A-layer + multi-year panel + closure / HCM anchor join). Higher-ed is at the "Phase 11 / 12 corporate" stage; one more higher-ed step adds more scientific evidence than one more corporate Phase-20 step. If a corporate Phase 20 is later required, the highest-marginal-value candidate is external divergence-anchor validation.
CORPORATE_PHASE19_DECISION_NOTE.md in the
project repository.
Back to Validation · Evidence · Documents index.