Improve AI Feature

Operation

Implement the requested change and verify its directly affected behavior.

Goal

Implement the selected AI feature change while preserving output validation, authorization, bounded execution and truthful user states.

Domain rules

• Validate and normalize model output at a typed boundary before side effects or user-visible state.
• Keep authorization, tenant filters, quotas, irreversible actions, and product decisions outside model control.
• Bound attempts, tool calls, context, output size, latency, and spend; expose truthful pending/failure/fallback states.
• Version prompts/config/artifacts with affected evals, caches, docs, and rollout controls.

Validation

• Run deterministic unit/integration/eval fixtures plus malformed output, timeout, refusal, fallback, auth, and schema cases.
• Repeat stochastic evaluations enough to report variance rather than one lucky sample.
• Record live-provider/model-version, safety review, cost, and production telemetry gates separately.

First reproduce the failed boundary using the smallest safe case. Correct the owning parser, prompt, context builder, policy or product-state transition, then replay the same input and its relevant refusal/error control. Compare semantic changes on the same evaluation inputs rather than a newly favourable example. A failed case returns to that owner; supported acceptance proceeds to Patch Checking on the full change.