Choose the action that preserves the task boundary and its evidence. These cases test template data, model settings, streaming, retrieval versions, release checks, and human review.
Review the decisions
- Prompt templates: bind variables without changing instruction structure
- Instruction conflicts: resolve authority before wording
- Model settings: change one generation variable against a fixed case set
- Streaming responses: validate the complete artifact before an effect
- Retrieved evidence: reconcile versions and conflicting facts
- Prompt caching: reuse stable context with a versioned expiry
- Metamorphic tests: verify behavior when harmless details change
- Human handoff: preserve evidence and the reason for uncertainty
- Prompt traces: connect an answer to its inputs, checks, and effects
- Model migration: replay contracts before changing providers or versions
- Fairness checks: test equivalent cases across groups and wording
Common Mistakes
- Do not accept a partial response as a final decision.
- Do not treat cached or retrieved text as current authority without version checks.
- Do not release from one attractive sample instead of a held-out case set.