RFC: Formalizing Epistemic Warrant Ceilings for Survival Analysis and Clinical Risk Scores

[HERRY423]

Motivation

As AI agents increasingly assist with biomedical and clinical data analysis, we need formal epistemic guardrails against small-sample survival analysis overclaims.

Proposed Rules for clinical.survival_analysis:

1. Event Count Boundary: If total uncensored events in any stratum $E < 10$, hazard ratio p-values must be clamped to FRAGILE maturity.
2. Proportional Hazards Assumption: Schoenfeld residual test failure ($p < 0.05$) must block single-hazard ratio reporting and require time-varying coefficient models.
3. Causal Blocking: Observational EHR survival curves must strictly block causal language (drug X improved survival $\rightarrow$ drug X was associated with longer overall survival).

Please share your thoughts on these proposed invariants before we formalize them into the BioNexus registry.