Commit the full decision record into the chain, not only the request

[President]

A ContextEntry commits only the request (request_hash, step_type, provenance_summary, timestamp).
Not the disposition, the rule(s) that fired, or the inputs it relied on (counters, classifier/model outputs).
So a Guardian can rewrite what it decided, and why, without breaking its own chain.

Suggest:

• each entry also commits a hash of the full decision record: request, disposition, reasons, inputs used, response.
  Whether the Guardian stores the record itself stays its choice/config; the hash is always in the chain.

[stevenzhao07]

Agree with the direction - a ContextEntry that commits only the request hash lets a Guardian rewrite "what and why" without breaking its own chain.

A few contributions from a working implementation of this exact pattern:

1. Reference implementation. ai-a3-laws ships a runtime guard that emits a complete decision record per decision: gate verdicts (A1 authority-scope, A2 human-oversight, A3 accountability identity), the inputs each gate relied on (authority_scope, human_review, actor_id, decision_id), and a PASS/BLOCK disposition with exit codes. The record is JSON and hashable, so it can feed exactly the chain-commit you describe.

2. What the committed hash should cover. Our practice has been an evidence triplet: input_hash (the decision payload the gates evaluated), output_hash (the disposition + reasons), and env_fingerprint (runtime/version). Committing a hash over request + disposition + reasons + inputs used + response - as you propose - plus the gate verdicts themselves, gives incident responders enough to reconstruct why without trusting the Guardian's own storage.

3. Complementary hook, not a replacement. Chain-commit (your #115) guarantees auditability after the fact. Separately, a "release gate" hook that must record pre-execution evidence before the tool call is released guarantees the evidence exists at decision time - otherwise a fast-failing harness can release a call with no record at all. The two together cover both ends of the decision lifecycle. Happy to draft the decision-record schema and a worked example mapping ai-a3-laws records onto the ACS trace vocabulary if useful.

[President]

@stevenzhao07 - thanks, agree! Will wait for others’ views on this.