← Mandate Registry · 7 mandate layers · model_collapse
tech_data · EU AI Act Article 10 · FTC AI guidance · CAC Generative AI Measures
Model Collapse / Verified Data
Model collapse from synthetic data — prove training data provenance or copyright defeat
—CAC algorithm filing + PIPL
Regulatory exposure — commonly overlooked:
Model trained on AI-generated data — vendor owns collapse risk, deployer owns downstream harm. · Algorithm filing requires decision logic documentation — independent verify acceptable to CAC review.
NYT litigation
Prove what data model saw — receipt at training decision point.
Art 10 EU
Training data governance mandatory — provenance receipt satisfies documentation.
Quality degradation
Collapse detection requires baseline receipt — compare model version decisions.
[China] Dual compliance
Multinationals operating CN + US/EU — one architecture, jurisdiction-tagged receipts.
[China] DeepSeek moment
Chinese model deployment globally — provenance receipt links model to decision.
[Text / Chat] Modality hook
Baseline — all frameworks apply to text decisions.
7 mandate layers (live)
Regulatory Clock
Countdown to operative regulatory deadline — NAIC adoption, GSE mandate, EU transposition.
Open →
Domain Classifier
Industry-specific SAFE/CRISIS/VIOLATION with regulatory framework mapping.
Open →
Exhibit / Filing Pack
Regulator-ready external validation — NAIC Exhibit D, Fannie QC, EU FRIA, FDA Part 11.
Open →
Mandate Registry
Enroll deployers/insureds under vertical-specific governance mandate.
Open →
Bind / Action Gate
ALLOW/BLOCK before consequential action — bind policy, sell loan, deactivate worker.
Open →
Compliance Certificate
Deployer-facing downloadable cert — forward to counsel, auditor, regulator.
Open →
Actuarial / Risk Feed
Anonymized asymmetry signals for pricing, reserving, reinsurance correlation.
Open →