Complete the governance instrument family: DPP, IAE, attribution-FAR

[HarperZ9]

What

Three re-derivable instruments under flywheel gov, each shipping its does-not-prove line and covered by unit + CLI tests.

• DPP (dataset_provenance.py, gov dpp) — verify a dataset's declared shard hashes and origins against the shards an evaluation actually observed; flag hash mismatch, undeclared shards, unknown-origin shards, and duplicate content. MATCH / DRIFT / UNVERIFIABLE. Motivated by the "distilled frontier data of unknown origin" gap and the local-first enterprise data-as-asset market.
• IAE (intervention_ablation.py, gov iae) — the paired-ablation test (an alignment intervention vs its ablated counterpart, misalignment judged per trajectory by a separate check); reports the misalignment-rate effect and a REDUCES / NO_EFFECT / INCREASES verdict, so a backfiring or inert intervention is a finding, not a silence. This is Daniel Tan's experiment as an instrument.
• attribution-FAR (attribution_far.py, gov far) — check the checker: measure how often an attribution method accepts a known-null control (false accept). RELIABLE / UNRELIABLE over the control corpus. The false-accept-corpus discipline applied to attribution.

Why

These complete the model-level instrument family (EIM, CED, NCEC, ERG, IGAP, SMP, ICP + these three). They keep the shared discipline: deterministic given the input, a verdict lattice, honest does-not-prove bounds, and a false-success control (compliance/misalignment/attribution judged by a separate signal, never self-confirming). attribution-FAR in particular makes the family's own attribution claims checkable — robust to the conflict-of-interest and capture concerns that motivate independent, re-derivable evaluation.

Tests

tests/test_dataset_provenance.py, tests/test_intervention_ablation.py, tests/test_attribution_far.py: verdict paths, computations, both-condition / min-control requirements, malformed inputs, threshold ranges, and the gov dpp|iae|far CLI (exit codes + dispatch through cli_entry). Full slice green; file-size gate clean; writing lint clean.

🤖 Generated with Claude Code