Claim Boundary Note

This note records what may and may not be claimed about the higher-ed extension. It is the canonical reference; any external restatement should preserve these boundaries.

Allowed claims

• The higher-ed extension has a hardened, narrow validated result for FASB × Sector 2 private nonprofit four-year institutions.
• The validated geometry is joint-weak G+A; in this lane it predicts closure within five years at AUROC ≈ 0.864.
• The result survives institution-cluster bootstrap uncertainty; cluster-bootstrap 95 % CI ≈ [0.806, 0.914].
• The result survives four sliding temporal folds at AUROC 0.77–0.86.
• The geometry has a higher AUROC point estimate than every tested single-variable baseline. Cluster-bootstrap paired-difference 95 % CI vs the strongest baseline is [+0.002, +0.138]; about 98 % of bootstrap samples are positive. This is a higher point estimate, not a definitive statistical-superiority claim.
• The closure signal is detectable across T-1 through T-5; the geometry produces a multi-year early-warning pattern, not a coincident T-1 alarm.
• The federal Financial Responsibility composite is itself a real closure-relevant signal on the descriptive train cohort 2014–2016 (AUROC ≈ 0.747).

Forbidden claims

• The extension is a college-failure prediction product. It is not.
• The extension works across all higher education. It does not.
• The extension works for public, for-profit, two-year, or community-college institutions. Those lanes are not validated.
• The geometry beats baselines for enrollment-collapse targets. It does not.
• Divergence (Appearance minus Ground) is the higher-ed signal. Empirically it is the weakest predictor in this lane.
• The work imports corporate co-collapse vocabulary as a positive claim. It does not.
• A single-scalar composite (such as Appearance over Ground) validates the higher-ed extension. No such scalar is computed.
• The extension is production-ready, suitable for automated decisions, an accreditation tool, an admissions tool, investment advice, a credit rating, or an institutional ranking. It is none of those.
• The extension establishes that the A/G geometry beats the federal Financial Responsibility composite on the locked test cohort. That comparison is not testable in this run.
• Joint-weak gives perfect early warning. Top-decile flagging captures only about 13–27 % of pre-closure rows across the T-1 through T-5 horizon.

Boundary checklist (must accompany any external restatement)

1. State the validated stratum (FASB × Sector 2 private nonprofit four-year).
2. State the validated target (closure within five years).
3. State the AUROC and the cluster-bootstrap 95 % CI ([0.806, 0.914]).
4. Mention that the original analytic Hanley-McNeil CI is [0.801, 0.928] for compatibility.
5. Use the language “higher AUROC point estimate over tested single-variable baselines,” not “statistically superior.”
6. Quote the temporal-folds AUROC range (0.77–0.86) and note the Fold-1 baseline tie.
7. Note that enrollment-collapse targets are baseline-dominated.
8. State that public, for-profit, two-year, and community-college applicability remain unvalidated.
9. State that the head-to-head comparison against the federal Financial Responsibility composite was not runnable on the locked test cohort.
10. State that HCM is not part of the validated evidence in this run.
11. State that the work uses a two-layer geometry; no single-scalar composite is computed.
12. State that the dominant closure regime is joint weakness (not divergence).
13. State that the result is research-grade, not production.

Project domain hierarchy

For context only:

• Banking — flagship operational dashboard / product (separate work).
• Higher-Ed FASB × Sector 2 closure — this hardened narrow validated branch.
• Corporate Filings — separate research-grade extension, locked at its own freeze.
• Other cross-domain tests — exploratory.

← Back to Evidence