evidence

Identity contracts for post-MIP evaluation evidence.

Functions

checkpoint_fingerprint	Return the checkpoint fingerprint used by evaluation identities.
collect_kd_exposure	Combine configured exposure with durable token and GPU-hour evidence.
downstream_evaluation_identity	Bind one comparison to its checkpoints, training, and evaluator-owned evidence.
evaluation_contract	Validate and project the generic fields that must match across evaluations.
evaluator_result_contract	Return the evaluator-owned comparison contract declared by one result.
exact_checkpoint_evidence	Read root-independent physical geometry and content identity from a checkpoint.
kd_exposure_metrics	Publish the comparable numeric exposure fields as node metrics.

checkpoint_fingerprint(checkpoint)

Return the checkpoint fingerprint used by evaluation identities.

Parameters:

checkpoint (str | Path)

Return type:

str

collect_kd_exposure(output, configured, *, max_steps, elapsed_gpu_hours, resumed_completed_milestone)

Combine configured exposure with durable token and GPU-hour evidence.

Parameters:

• output (Path)

• configured (Mapping[str, Any])

• max_steps (int)

• elapsed_gpu_hours (float)

• resumed_completed_milestone (bool)

Return type:

dict[str, Any]

downstream_evaluation_identity(*, source, reference_checkpoint, profile, evaluator_revision, candidate, reference, reference_checkpoint_fingerprint=None)

Bind one comparison to its checkpoints, training, and evaluator-owned evidence.

Parameters:

• source (CandidateRevision)

• reference_checkpoint (str | Path)

• profile (Any)

• evaluator_revision (Any)

• candidate (Mapping[str, Any])

• reference (Mapping[str, Any])

• reference_checkpoint_fingerprint (str | None)

Return type:

dict[str, Any]

evaluation_contract(identity, *, label, expected_profile, expected_reference_fingerprint)

Validate and project the generic fields that must match across evaluations.

Parameters:

• identity (Any)

• label (str)

• expected_profile (str)

• expected_reference_fingerprint (str)

Return type:

dict[str, Any]

evaluator_result_contract(result)

Return the evaluator-owned comparison contract declared by one result.

Parameters:

result (Mapping[str, Any])

Return type:

dict[str, Any] | None

exact_checkpoint_evidence(checkpoint)

Read root-independent physical geometry and content identity from a checkpoint.

Parameters:

checkpoint (str | Path)

Return type:

dict[str, Any]

kd_exposure_metrics(exposure)

Publish the comparable numeric exposure fields as node metrics.

Parameters:

exposure (Mapping[str, Any])

Return type:

dict[str, float]