Validation summary
Phase 18 full-population pooled metrics
Threshold-fixed at z = 0.0 (no test-fold tuning). 500-resample CIK-cluster bootstrap CIs.
Strict label (direct + high/medium-confidence anchors)
| Geometry | Internal metric | AUROC | 95% CI | AP | Lift @ 10% | Lift @ 1% |
|---|---|---|---|---|---|---|
| Structural distress | structural_distress_lam0_0 | 0.711 | [0.660, 0.753] | 0.023 | 2.21× | 0.43× |
| Joint-collapse severity | co_collapse_lam0_0 | 0.677 | [0.631, 0.726] | 0.021 | 1.90× | 0.00× |
| Divergence / masking | neg_R_lam0_0 | 0.489 | [0.419, 0.551] | 0.016 | 2.03× | 3.03× |
All-anchored label (every Phase-14 anchored CIK)
| Geometry | Internal metric | AUROC | 95% CI | AP | Lift @ 10% | Lift @ 1% |
|---|---|---|---|---|---|---|
| Structural distress | structural_distress_lam0_0 | 0.683 | [0.639, 0.719] | 0.030 | 1.81× | 0.85× |
| Joint-collapse severity | co_collapse_lam0_0 | 0.642 | [0.602, 0.683] | 0.027 | 1.55× | 0.28× |
| Divergence / masking | neg_R_lam0_0 | 0.510 | [0.457, 0.554] | 0.024 | 2.03× | 2.54× |
The candidate geometries’ CIs cleanly exclude the
0.5 chance line on the lower side. neg_R
brackets 0.5 — chance, as expected for a divergence /
masking diagnostic on a broad Chapter 11 / Item 1.03
distress screen.
Phase chain
The corporate signal has been measured across five successive cohort designs. The story is one of real but attenuating discrimination as the cohort moves from enriched (engineered) to realistic (full population).
| Phase | Cohort | Test base rate | Structural AUROC | Joint-collapse AUROC | Divergence AUROC |
|---|---|---|---|---|---|
| 15b | Enriched held-out | 0.71 | 0.72 | 0.74 | 0.39 (worse than chance) |
| 16 | Hardening / base-rate | 0.19 | 0.71 | 0.66 | 0.49 |
| 17 | Population sample (708 CIKs) | 0.07 / 0.11 | 0.78 / 0.76 | 0.76 / 0.73 | 0.45 / 0.47 |
| 18 | Full population (3,412 CIKs) | 0.012 / 0.018 | 0.711 / 0.683 | 0.677 / 0.642 | 0.489 / 0.510 |
| 19 | Hardening / freeze | — | (unchanged) | (unchanged) | (unchanged) |
Slashes separate strict / all-anchored where both are reported. The geometry ordering (structural ≥ joint-collapse > divergence) is preserved at every cell. Phase 18 is the locked final result.
Baseline comparison (Phase 19)
Two simple primitives — neg_g4_ocf_assets_z
and neg_g2_interest_coverage_z — slightly
outperform structural_distress_lam0_0 on AUROC.
This weakens any claim that the A/G composite is the
strongest raw discriminator, but it does not erase the
structural / regime interpretation value.
| Score | Kind | AUROC (strict) | Lift @ 10% |
|---|---|---|---|
| Operating cash flow / assets (risk-up) | baseline | 0.736 | 2.34× |
| Interest coverage (risk-up) | baseline | 0.726 | 1.93× |
| Altman-like proxy (Phase-19 weighted; not actual Altman Z) | baseline (proxy) | 0.714 | 1.94× |
| Return on assets (risk-up) | baseline | 0.713 | 1.82× |
| Structural distress (composite) | candidate | 0.711 | 2.21× |
| Joint-collapse severity (composite) | candidate | 0.677 | 1.90× |
The composite’s value is robustness (graceful degradation when a primitive is missing) and regime separation, not maximum raw discrimination. Full table: Baseline Comparison.
Sector stratification (Phase 19)
Pooled AUROC of 0.711 is broadly distributed across reliable SIC divisions (n_pos ≥ 20 strict positives in cell):
- Wholesale + Retail Trade — AUROC 0.742, n_pos = 30. Strongest reliable cell.
- Manufacturing (food / textiles / chem / petroleum) — AUROC 0.729, n_pos = 85. Largest contributor.
- Manufacturing (durables / electronics) — AUROC 0.639, n_pos = 53. The weak cell.
Smaller cells (Services, Health/Education, Mining, Transport, Agriculture, Finance/RE residual) have fewer than 20 strict positives and are flagged DESCRIPTIVE_ONLY or NO_STRONG_CLAIM. Their flashy AUROCs (e.g., Finance/RE residual 0.92 on 6 positives) should not be over-interpreted. Full table: Sector Analysis.
Calibration / monitoring bands (Phase 19)
Predeclared (not tuned) quantile bands on the test-fold score distribution for structural distress:
| Band | Percentile | n in band | Pos in band | Event rate | Lift | Recall captured |
|---|---|---|---|---|---|---|
| Normal / background | 0–60 | 11,646 | 49 | 0.42% | 0.35 | 21% |
| Elevated | 60–80 | 3,882 | 76 | 1.96% | 1.65 | 33% |
| Watch | 80–95 | 2,911 | 90 | 3.09% | 2.60 | 39% |
| High structural concern | 95–99 | 745 | 15 | 2.01% | 1.69 | 6.5% |
| Critical / inspect manually | 99–100 | 226 | 1 | 0.44% | 0.37 | 0.4% |
Tail-dropoff notice. The Critical / top-1 % band does not have the highest event rate. Its lift is below 1.0 (worse than random for that slice). Treat it as noise-prone, not as the highest-priority flag. The most informative tier is the Watch band (80–95th percentile) — 39 % of strict positives captured by 15 % of the universe, lift 2.6×.
These bands are research-grade triage primitives, not automated decisions. Each company appearing in any band requires manual inspection. Full policy: Calibration Policy.
Divergence-anchor study (Phase 19)
Inside the divergence-state cohort (firm-years where A > 0 and G < 0 — the “healthy-looking earnings, weak balance-sheet ground” pattern), the divergence / masking geometry shows the best discrimination of future structural deterioration:
| Geometry | Divergence-state cohort AUROC | All-comers cohort AUROC |
|---|---|---|
| Structural distress | 0.666 | 0.465 (mean reversion) |
| Joint-collapse severity | 0.468 (chance) | 0.552 |
| Divergence / masking | 0.705 | 0.342 (mean reversion) |
The 36-percentage-point AUROC swing for the divergence geometry between cohorts (0.342 broad → 0.705 inside divergence-state) is first internal evidence that the canonical R / divergence geometry behaves as the Phase-13 doctrine predicted — useful for the divergence / masking regime, not for broad Ch.11 screening. This is an internal future-state anchor, not an external divergence-event anchor; external validation is a future-phase candidate. Full study: Divergence-Anchor Study.
Anchor refresh (Phase 19)
A bounded EDGAR full-text-search refresh (2026-01-01 → 2026-04-26) found 6 new high-confidence Item-1.03 anchors; 5 of 6 are already in the Phase-18 panel. The descriptive out-of-time check (22 pre-petition firm-year rows) is consistent with Phase-18 in direction but is too small for a statistical claim. Full note: Anchor Refresh Note.
Bottom line
- The full-population signal survives with documented attenuation.
- Structural distress is the strongest A/G composite screen.
- Two simple primitives slightly beat the composite on AUROC; the composite’s value is robustness and regime separation, not maximum discrimination.
- Canonical R / divergence is at chance for broad Ch.11 screening; first internal evidence supports it for the divergence / masking regime.
- Corporate stands as a research-grade SEC public-company extension: not a production risk model, automated decision system, investment tool, credit rating, or replacement for credit analysis.