# Regulated AI Change-Control Pilot

CHANGE → E-DTL → E-DCLA → E-DELA → E-DEGA → exact operator review → E-ACC → E-RA → receipt / rollback / replay.

The synthetic AI_SYSTEM_A is a document-classification configuration, with model configuration, prompt, software component, control pack, evidence EV001/EV005/EV006, evaluation, current release, release candidate and evidence pack. No model inference runs. Dependencies in pilot/manifest.json define the exact test boundary.

Changed versions invalidate only dependent evidence. EV005 remains current during a prompt change; the evaluation and release conclusions become untrusted through EV001. Validity is checked against an explicit recorded run time. Same frozen input, state and time yield the same declared result; receipt IDs, wall-clock timestamps and measured durations vary.

E-DEGA PASS permits review, not an effect. E-ACC checks the actor and exact review digest. E-RA verifies Ed25519, expiration, one-use state, current actor/target/scope, before-state, candidate, policy, graph, evidence, dependency versions, review and rollback before committing. Cases deliberately change these inputs between issuance and commit. Every effect stays inside the audit workspace.

E-OSA and E-VSC explain the research beneath the wider system. This pilot does not require operating them, trusting an AI-generated regulatory interpretation or adopting the whole research stack. E-SR is explicitly absent.

Measured component timings exclude authentication, network, human review, UI, evidence signing and export; export and signing have separate records. No end-to-end sub-millisecond promise is made.