AI CHANGE LOG STARTER KIT

Goal
Make AI behavior changes explainable and reversible. This is not a policy document. It’s an operating artifact that produces a human-readable log entry for every meaningful AI change.

1) PR TEMPLATE (required for any AI behavior change)
Copy into every PR that changes: model selection, prompts/system messages, retrieval sources, tool permissions, routing logic, or safety rules.

- Change type: Model | Prompt | RAG corpus | Tooling | Policy | Routing
- Summary (1–2 lines):
- User-facing behavioral diff (include 2–3 before/after examples):
- Dependencies touched (be specific):
 - Model name/version:
 - Prompt/template identifier (hash or file path):
 - Retrieval corpus/index ID and source systems:
 - Tool scopes/permissions changed:
- Risk assessment (what can go wrong; who is impacted):
- Evals/tests run (links):
 - Offline eval suite:
 - Adversarial/red-team notes (if relevant):
- Rollout plan:
 - Dev verification steps:
 - Canary plan (who/what gets it first):
 - Monitoring signals (what you’ll watch):
- Rollback plan (exact steps):
 - How to revert prompt/model/routing:
 - How to invalidate caches/rebuild index (if needed):
- Ownership:
 - DRI (named person/team alias):
 - Reviewer (named):
 - On-call contact for incidents:

2) AI RELEASE NOTE FORMAT (human-readable)
Publish each merged change in a place non-engineers can see (internal wiki page or a dedicated Slack channel).

- Date:
- What changed:
- Who approved it:
- Who owns it (DRI):
- Who it affects (teams/users):
- What to watch for (symptoms of regression):
- How to report issues (link to ticket form / Slack channel):
- Rollback status:
 - “Reversible in minutes” / “Reversible with rebuild” / “Not reversible (blocked) — explain why”

3) MONTHLY ROLLBACK DRILL (30 minutes)
Run this like a lightweight fire drill.

Step 1: Pick one recent AI change (prompt, model, corpus, tool scope).
Step 2: Simulate a regression report with a concrete example.
Step 3: Confirm you can answer, from the log alone:
 - What changed?
 - When did it ship?
 - Who approved it?
 - What dependencies are involved?
Step 4: Execute rollback steps in a non-prod environment (or via a feature flag) and confirm behavior returns.
Step 5: Write one improvement to the template/runbook based on friction encountered.

Non-negotiables
- No unowned AI in production: if nobody is on-call, it doesn’t ship.
- Every AI change must be reversible or explicitly blocked with a written rationale.
- If a change cannot be explained in plain language, it isn’t ready to deploy.

If you adopt only one thing from this kit: require a rollback plan in every AI change PR. It forces engineering discipline, and it forces leadership clarity about risk.