Answers and explanations · Preserve the experiment you actually ran

1. An unchanged freshly generated manifest verifies with []. Replacing the saved policy digest with zeros returns ['policy.json']: a named artifact differs from the recorded claim. Removing that name makes the required set incomplete and raises InputError. Neither result proves a model changed.
2. System: changing the prompt requires generating authorized comparable outputs and re-evaluating. Dataset: correcting a label requires re-scoring both versions on the corrected version and keeping the old result. Grader: revising grounded requires re-grading both outputs and preserving grader identity. Policy: changing a latency ceiling requires applying the reviewed policy to the same relevant metrics and recording a new decision, rather than pretending the measured system improved.
3. A hash describes bytes. Formatting can change bytes without changing parsed values. Inspect the semantic diff, revalidate inputs, and replay the same evaluator. A real regression would need a behavior or evaluation difference under comparable conditions; a digest difference alone cannot establish it.
4. Example record: task and sampling unit; dataset version and provenance; system/model/prompt/configuration versions; authorized output IDs; grader/rubric version; policy version; interpreter/dependency context; start/end timing definitions; cost accounting convention; result digests; decision owner; limitations. A retention decision could keep only synthetic data now, and require a separately approved minimal-field policy before any real collection. There is no need to put an API key into an experiment record.

Self-check: your answer should distinguish reproducible scoring from reproducible generation, identity from authenticity, and a decision rule from an observation. These distinctions matter even when every program in the exercise is deterministic.