Annex D

Catastrophic-Risk Evaluation (CRE) Protocol


ANNEX D CATASTROPHIC‑RISK EVALUATION (CRE) PROTOCOL

D‑1 Trigger Criteria A system must pass a CRE before deployment if it meets any of the following criteria: (a) Training compute exceeds 10²⁶ FLOP. (b) Autonomous transactional authority averages > $10 M/day. (c) Recursive Event: Any initiation of autonomous code generation, weight modification, or hyper‑parameter tuning capabilities intended to alter the system's own cognitive architecture, objective functions, or PDMA logic.

D‑2 Required Artefacts

  1. Independent red‑team report (≥ 1 FTE‑month).
  2. Interpretability / latent‑goal probe study.
  3. Kill‑switch & containment test results.
  4. Comparative baseline vs. current frontier models.
  5. Dual sign‑off by two Wise Authorities outside the developing organisation.

D‑3 Publication & Escrow
• Summary report public within 30 days. • Full technical package escrowed with a recognised national safety authority.

D‑4 Re‑Certification
• Mandatory after any major model revision (> 2 % parameter delta or architecture change).

D‑5 Failure Response
• Deployment blocked until deficiencies remediated and re‑audited.

On this page