These cases ask which layer failed and what evidence is required before a result can be used. Select the answer that keeps uncertainty and downstream effects explicit.
Review the decisions
- Prompting, retrieval, or training: choose the failing layer
- Retrieval queries: resolve references without adding facts
- Document chunks: preserve the clause and its governing exception
- Numeric prompts: let code calculate and the model explain
- Selective answers: measure when to abstain
- Evaluation leakage: keep the release test independent
- Generated output: validate again at the destination boundary
Common Mistakes
- Do not blame prompt wording when the governing document was never retrieved.
- Do not count a fragment as a complete rule.
- Do not treat schema-valid text as safe at every destination.