# Snapshots / approved output Also: approval tests, golden files/masters, reference-output comparisons. Characterization describes a purpose, not necessarily snapshots. - Fit: consumer-visible formats, protocols, generated files/data, text whose representation matters. - Review initial baselines against intent, requirements, or independent examples. Recording current output ≠ validating it. - Keep output focused, readable, deterministic. Stabilize inputs; normalize only irrelevant variation. - Keep semantic checks where multiple representations are valid. Do not accidentally mandate irrelevant ordering, whitespace, or prose. - Snapshots describe selected examples, not the whole input domain. ## Failed comparison 1. Inspect the diff; identify why output changed. 2. Compare with authorized intent and consumer contracts. 3. Fix regressions; regenerate only justified intentional changes. 4. Review the complete affected baseline diff, including removals and normalization. 5. Rerun relevant comparisons and behavioral checks; disclose unresolved differences. No bulk refresh to silence failures. Generated ≠ approved. Follow existing approval requirements, not a new ceremony. During curation, replace repetitive field assertions with focused approved output where cases remain protected. Replace internal dumps with meaningful output or semantic checks. Keep targeted safety/protocol assertions when clearer. Appearance-dependent acceptance → use the visual-verification skill. Snapshot changes may invalidate mutation history. ## Sources - [ApprovalTests](https://approvaltests.com/) - [Jest: review and scoped updates](https://jestjs.io/docs/snapshot-testing)