What evidence makes an agent behavior change reviewable after observability flags it?

[vittorfp]

I am looking for practical feedback from people doing LLM/agent observability.

I am testing a small open-source artifact called change_proposal: a plain JSON file that sits downstream of traces, evals, and feedback. It records the observed behavior, evidence, proposed behavior change, validation criteria, risk, and rollback plan.

Repo: https://github.com/vittorfp/agent-change-proposals
Example: https://github.com/vittorfp/agent-change-proposals/blob/main/examples/declared-intent/change_proposal.example.json
Discussion: vittorfp/agent-change-proposals#6

This is not meant to replace observability tooling. The question is the handoff after observability finds a problem: how should a proposed prompt/tool/retrieval/guardrail change be made reviewable?

I would value feedback on:

1. Would a file like this fit PR review, eval review, CI, or incident follow-up?
2. What trace/eval evidence must be preserved for reviewers to trust the proposal?
3. Are fields like immutable evidence refs, baseline/candidate IDs, evaluator version, dataset version, metric threshold, sample count, and rollback trigger enough for CI gating?
4. Does this belong as plain JSON first, or should it map directly to an observability platform model early?