1. What this is not
- Not a bankruptcy prediction product.
- Not investment advice.
- Not a credit rating.
- Not a production risk model.
- Not a replacement for credit analysis.
- Not a production risk model or automated decision system.
2. Main limitations
Single primitives slightly outperform the composite on AUROC
On the Phase-18 test fold, two single-feature baselines —
neg_g4_ocf_assets_z (operating cash flow ÷
assets, oriented risk-up) AUROC 0.736 and
neg_g2_interest_coverage_z AUROC 0.726 —
slightly beat structural_distress_lam0_0
(AUROC 0.711) by 1.5–2.5 pp. The composite’s value
is robustness (graceful degradation when one primitive is
missing) and regime separation, not maximum raw
discrimination.
Low base rate limits precision and F1
Eligible-universe Chapter-11 incidence is on the order of 0.5 – 1.0% per year. Even an AUROC of 0.71 produces low precision at any practical operating threshold; the Watch-band precision is roughly 3%. F1 collapses at low base rate and is not used as a headline metric.
Top-1% tail dropoff
The top 1% of structural-distress scores has lower event rate than the Watch band (80–95th percentile). The very tail is noise-prone. Operators should not treat the top 1% as the highest-action band.
Sector heterogeneity remains
Pooled AUROC 0.711 hides cell-to-cell variation. The Manufacturing-3 (durables / electronics) sector cell shows AUROC 0.639 — meaningfully weaker than the pooled headline. Wholesale / Retail Trade and Manufacturing-2 (food / textiles / chem / petr) are the strongest reliable cells.
Item 1.03 anchor semantics imperfect
Even after the LLM-assisted scope classifier, about 25%
of anchored CIKs sit in ambiguous_item_103 or
non-direct buckets. The strict-label cohort (231
positives) is the cleanest cell; the all-anchored cohort
(354 positives) introduces some scope noise.
XBRL coverage / filtering matters
Companies with thin XBRL coverage are excluded by design. The Phase-12 SIC + facts-size + 3-Assets gates are conservative; some genuinely-distressed firms are filtered out before scoring.
Bankruptcy paths beyond Item 1.03
Chapter-7 cases, Item 5.02 / Item 8.01 attachments, and going-private transactions are not captured by the anchor table.
Divergence / masking — internal evidence only
The Phase-19 divergence-anchor study used an internal future-state anchor (G drops by ≥ 1 z over 1–2 years OR firm enters pre-petition state). The result is suggestive (canonical R AUROC 0.705 within the divergence-state cohort) but it is not external validation. An external divergence-anchor study is a future-phase candidate.
Current company rankings are not published
This site does not distribute per-company distress scores, top-N watchlists, or any ranked list. That is a deliberate constraint, not an oversight.
3. Why it is still worth keeping
- The full-population signal survives with documented attenuation across a five-cohort validation chain.
- The geometry roles are interpretable and separate broad structural distress, near-petition severity, and divergence / masking.
- The work demonstrates an A/G-style structural decomposition adapted to SEC public-company filings, with corporate-specific geometry and limitations.
- The chain provides a research artifact with public-data lineage, reproducible from raw SEC endpoints.
4. What would justify reopening corporate
The corporate branch is frozen as of Phase 19. A future Phase-20 step would only be justified if one of the following is true:
- Out-of-time anchor refresh — quarterly continuation of the EDGAR FTS query, broader item-code coverage (Item 5.02 / 8.01 attachments), and a partial Phase-18 panel rebuild for any new admitted CIKs.
- External divergence-anchor validation — convert the Phase-19 internal future-state anchor into an external one using restatements / going-concern / material-weakness disclosures.
- Sector-stratified expansion — deepen the analysis in the weak Manufacturing-3 cell to understand whether the durables / electronics signal is recoverable.
- Calibration audit — isotonic / Platt scaling on train, applied to test, replacing quantile bands with a calibrated probability output.
- Benchmark comparison against more established distress models (e.g., a properly-built Altman Z-score where the required variables are reliably available).
Discipline note
We chose to lock at Phase 19 with the baseline-comparison finding visible rather than to suppress it or tune it away. The goal of this site is to make the corporate evidence chain available with its limits intact, not to advertise a stronger conclusion than the data supports.