Release evidence, stress-tested

Will your model fingerprint survive the derivative?

Compare provenance strategies against the transformations you expect after release. Scores explain assumptions; they never claim legal proof.

Release profile

Set the operating conditions

Model family Release type Model size Verifier latency

Maximum utility loss 1.5%

False-positive tolerance 0.5%

Derivative attack simulator

Change what happens downstream

2 active

4-bit quantizationFormat + precision shift Fine-tuning5-15% task adaptation PruningRemove low-weight paths DistillationTransfer behavior to student Format conversionRepack and reshard weights Adaptive removalFingerprint-aware attacker

Balanced sample loaded. Toggle an attack to see the recommendation move.

Strategy comparison

See where each method breaks

Evidence bundle

Prepare the proof trail

3 of 6 ready

Verification plan

Prove a chain, not a single signal.

The selected release profile needs complementary evidence because transformations can erase or imitate any single fingerprint.

Next actionRun controlled quantization and fine-tuning tests, then archive verifier outputs with the signed release manifest.

Decision

Use the hybrid bundle for this public open-weights release.

Trust boundary

This plan organizes evidence. It does not establish legal ownership or guarantee detection.

Assumption discipline

Record what changed, what survived, who replicated it, and where the claim stops.

Reusable artifact

Leave with a release decision record.

Save a scenario for comparison, then copy or download the exact assumptions behind the recommendation.

Saved scenarios

0 saved

Save this release profile as a comparison baseline.

Provenance brief

Plain text + JSON