This site documents a research extension of the A/G framework to public SEC company filings. It tests whether structural-financial deterioration patterns can be separated into interpretable regimes across eligible non-financial operating companies. It is a research-grade evidence portal, not a service.
What the evidence says
Signal survives full-population testing
Structural distress remains the strongest A/G composite screen under full eligible-universe testing on 3,412 admitted CIKs.
The signal attenuates
Performance drops from enriched / sample settings to the harder full-population base-rate setting (about 1% per year).
Simple primitives matter
Operating cash flow / assets and interest coverage slightly outperform the composite on AUROC.
Value is structural interpretation
The A/G architecture separates broad structural distress, joint-collapse severity, and divergence / masking — not a single discrimination edge.
At a glance
- Dataset
- Eligible SEC non-financial operating companies under project filters
- Admitted CIKs
- 3,412
- Panel rows
- 38,748
- Post-gate rows
- 36,041
- Test rows
- 19,410 (FY 2019–2025)
- Strict / direct positives
- 231
- All-anchored positives
- 354
- Primary composite screen
- Structural distress
- Status
- Research-grade / not production
Main finding
The full-population signal survives with attenuation. Structural distress is the strongest A/G composite screen, but two simple financial primitives — operating cash flow / assets and interest coverage — slightly outperform it on AUROC. The corporate extension is best understood as a research-grade structural interpretation layer, not as a maximum-discrimination prediction tool.
Key validation table
Phase 18 full-population pooled metrics on the test fold (FY 2019–2025; 19,410 rows; 500-resample CIK-cluster bootstrap CIs).
| Geometry | Internal metric | AUROC (strict) | AUROC (anchored) | Role |
|---|---|---|---|---|
| Structural distress | structural_distress_lam0_0 | 0.711 | 0.683 | Primary A/G composite screen |
| Joint-collapse severity | co_collapse_lam0_0 | 0.677 | 0.642 | Near-petition / joint-deterioration geometry |
| Divergence / masking | neg_R_lam0_0 | 0.489 | 0.510 | Not a broad Chapter 11 / Item 1.03 distress screen; retained for divergence regime |
Pooled base rate 1.19% strict / 1.82% all-anchored. CIs for both candidate
geometries cleanly exclude the 0.5 chance line; neg_R brackets 0.5,
consistent with its role as a divergence diagnostic rather than a Ch.11 screen.
Claim boundary (compact)
Corporate A/G is a research-grade SEC public-company structural-distress extension. It is not a bankruptcy prediction product, not a credit rating, not investment advice, and not a replacement for credit analysis. Simple primitives slightly outperform the composite on broad Chapter 11 AUROC, so the corporate composite should not be described as the strongest binary predictor; its supported value is diagnostic — separating structural distress, joint-collapse severity, and divergence / masking regimes. The full discipline is on the Limitations page; the operator’s public claim discipline lives in Final Freeze Note.
Where to go next
- Methodology — data sources, universe construction, geometry definitions, why corporate requires its own geometry.
- Validation — full-population pooled metrics, comparison chain, baseline comparison, calibration policy, divergence study.
- Evidence — rendered summaries of the source documents.
- Limitations — the blunt list of what this is not.
- Commercial Use — access boundary and contact.
Related Public Demos
Related public demos from the same project family. Each demo has its own scope, assumptions, validation status, and limitation notes.
A/G Bank Scoreboard
Public dashboard for A/G-style bank risk monitoring using appearance/ground divergence signals and documented validation boundaries.
A/G Higher Education
Exploratory A/G-style higher-education institutional risk/stability demo with explicit data-scope limits.
Bounded Replay-Risk Audit
Bounded workflow replay-risk demo that checks whether scoped AI-agent/SaaS workflows can reach prohibited states within a configured model and bound.
Scope and assumptions differ by demo.