A07 — Certified inputs and artifact lineage

• Input [direct]: Experiment specification, provider/request configuration, raw response, normalized result, source digests, and run metadata.
• Operation [direct/inference]: Assign stable artifact and run identifiers; snapshot the request; retain raw and normalized artifacts; record source and request digests; certify dataset completeness before comparison; keep approved instruments frozen unless a blocking defect requires a documented versioned change.
• Output [direct]: Canonical run directory, normalized result, normalization certificate, metadata, and downstream comparison inputs.
• Scope conditions [direct]: Evidence runs and experiments where reproducibility and dataset completeness matter.
• Failure conditions [direct]: Missing, blocked, invalidated, or stale certificate; comparison against uncertified data; unrecorded change to a frozen instrument.
• Predicted effect [absent]: No controlled directional outcome is stated for this mechanism.
• Measurement [direct/inference]: Certificate state, schema validity, digest agreement, dataset completeness, and presence of required artifacts.
• Named baseline [absent]: No matched control is named.
• Boundary condition [direct]: The rules govern the research pipeline; a documented blocking defect permits a versioned emergency change.
• Implementation test [direct]: Attempt official comparison with a missing or stale certificate; the pipeline must refuse it.
• Decision rights [direct]: Approved instrument remains frozen; accepted evidence decision records govern downstream experiment selection.
• Uncertainty handling [inference]: Missing or invalid lineage remains a blocking state rather than being silently repaired.
• Verification [direct]: Normalize, verify completeness, generate certificates, and compare only certified datasets.
• Provenance [direct]: Stable IDs, timestamps, provider/model data, request/result paths, and SHA-256 fields in schemas and run layout.
• Implementation cost [inference]: Snapshotting, normalization, certification, hashing, and retained storage.
• Limitations [direct/inference]: Repository-wide append history and supersession behavior are not fully defined by the inspected canonical records.
• Fidelity check [direct]: Official comparison refuses non-current certificates and consumes only canonical normalized paths.
• Null-capable future experiment [inference]: A reconstruction audit can show no improvement in reproducibility or can expose unresolved lineage gaps.

Statement basis

• Direct: Freeze rule, certification gate, schemas, identifiers, digests, verification sequence, and refusal behavior.
• Inference: Consolidated operation, uncertainty behavior, cost, broad limitation, and future experiment.
• Proposal-only: None.
• Absent: Controlled effect and named baseline; comprehensive supersession policy.