AI Model Change Safety & Compatibility Review

The likely issue is a contract mismatch around AI model/version changes, capability drift, structured outputs, tool use, latency, cost, safety policy, prompt compatibility, evaluations, rollout, rollback, and evidence gates. I would first define the invariant: A model change is a behavioral dependency change and must not reach full rollout until critical product, policy, tool, format, latency/cost, and rollback requirements are evidenced at an agreed threshold. Then I would trace the smallest owning boundary, apply or propose a bounded repair/review step, and add deterministic normal, boundary, failure and replay tests. The workflow preserves existing product requirements, safety policy, privacy boundaries, customer commitments, and release governance. If repository/workspace files are available, the findings can be mapped to exact paths and focused local validation; with partial evidence, the result stays an explicit diagnosis/evidence map with assumptions and no invented execution claims.