Data, universe, and geometry

1. Data sources

Every input comes from the U.S. Securities and Exchange Commission’s public EDGAR system. There are no paywalled or proprietary feeds, no Compustat / WRDS, and no third-party news.

• SEC ticker master — data.sec.gov/files/company_tickers.json: the public list of registrant CIKs and tickers.
• Submissions metadata — data.sec.gov/submissions/CIK*.json: per-registrant SIC code, fiscal year end, and filing list.
• Companyfacts (XBRL) — data.sec.gov/api/xbrl/companyfacts/CIK*.json: per-registrant XBRL fact tables.
• Chapter 11 / Item 1.03 distress anchor layer — 8-K filings whose structured items field contains 1.03 (Bankruptcy or Receivership), found via SEC EDGAR full-text search; 8-K body text fetched for scope classification.

Polite SEC pacing (default 0.18 s between requests, well under SEC’s 10/s cap), declared User-Agent, atomic per-file writes. Provenance is recorded in the project repository.

2. Universe construction

Non-financial operating-company filters

• Exclude banks (SIC 6000–6299).
• Exclude insurers (SIC 6300–6499).
• Exclude SPACs / funds (SIC 6770 / 6726).
• Exclude registrants with missing or non-numeric SIC.

XBRL coverage gates

• Companyfacts JSON ≥ 500 KB (drops shells and very-small filers).
• ≥ 3 distinct annual Assets rows from 10-K / 10-K/A filings.
• Per firm-year, at least 3 of 4 G primitives present (g_coverage_gate ≥ 0.75).

Train / test split

• Train years: FY ≤ 2018.
• Test years: FY 2019 – FY 2025.
• The split rule has been preserved unchanged from Phase 15b through Phase 19 so the comparison chain is valid.

After all four gates, the locked Phase-18 panel admits 3,412 distinct CIKs; the post-coverage-gate panel is 36,041 firm-year rows.

3. G and A in corporate filings

The A/G framework separates two layers of a firm’s financial state:

G primitives — structural condition (“ground”)

• Current ratio (g1_current_ratio) — current assets ÷ current liabilities.
• Interest coverage (g2_interest_coverage) — operating income ÷ interest expense.
• Balance-sheet cushion (g3_bs_cushion) — cash and equity buffer relative to liabilities.
• Operating cash flow / assets (g4_ocf_assets).

A primitives — operating / performed appearance

• Return on assets (a1_roa).
• Gross margin (a2_gross_margin).
• Year-over-year revenue growth (a3_rev_growth).

G is the structural ground; A is the operating surface. The Phase-13 doctrine is that distress regimes can be separated by how G and A move together (joint collapse) or apart (divergence / masking).

Corporate uses state-vector geometry: the geometries below are operations on standardized G and A aggregates, not on a single composite scalar. This keeps the structural-vs-operating decomposition explicit and lets different distress regimes (joint collapse, divergence / masking) be separated rather than collapsed into one number.

Transforms (train-only fit)

• Winsorize each primitive at the 5th / 95th percentile, fit on train years only.
• Robust z-standardize: (x − train median) / (1.4826 × train MAD).
• Apply fitted transforms to test rows. A pre-execution assertion enforces zero index overlap between train and test.

4. Geometry definitions

Three risk-up geometries are computed at λ = 0 (the Phase-13 / Phase-15 corporate primary):

Structural distress

Plain language: “The firm’s structural balance-sheet ground is weak.” A higher score means weaker G.

Formula: structural_distress = −G_score (the G aggregate, oriented so higher = more distress).

Joint-collapse severity

Plain language: “Both the structural ground and the operating surface are weak together.” A higher score means joint collapse and is informative for near-petition severity, not broad screening.

Formula: co_collapse = −(G_score + A_score).

Divergence / masking

Plain language: “The operating surface looks healthy while the structural ground is weak.” A higher score means more divergence between A and G. This is a diagnostic, not a Ch.11 screen.

Formula: neg_R = A_score − G_score (= the canonical R viewed risk-up).

A diagnostic weakest_channel_fragility (binding-constraint indicator) is also computed but is retained as descriptive only — its persistence-smoothed validation is deferred to a future phase.

5. Why corporate requires its own geometry

Corporate SEC/XBRL analysis has constraints that shape what this extension can and cannot claim. These constraints define the corporate claim boundary; they are not deficiencies relative to any other domain.

• SEC / XBRL data is heterogeneous. Tag variation across filers, restatements, late filings, and small-cap shells all add noise that the geometry has to absorb.
• Chapter 11 / Item 1.03 distress is a low-base-rate event, on the order of 0.5 – 1.0% per year for the eligible public-company universe. This caps achievable precision at any practical operating threshold.
• Sector heterogeneity is large. Manufacturing, Wholesale/Retail, and Services have different distress patterns and disclosure conventions, so pooled metrics hide cell-to-cell variation.
• Anchor coverage is imperfect. Corporate uses Item 1.03 8-Ks only — missing some Chapter-7 paths and bankruptcy events filed under Items 5.02 / 8.01 attachments.
• State-vector geometry is the right shape for these constraints. Keeping G and A separate lets the distinct regimes (joint collapse vs divergence / masking) be diagnosed rather than averaged out into a single risk score.

6. What was not done

The full reproducibility runbook lives inside the project repository (not duplicated here). Independent reproduction would require re-fetching from the public SEC EDGAR endpoints; raw JSONs are not distributed from this site.