1 lightning.finite_horizon.diagnostics.population

lightning.finite_horizon.diagnostics.population

Shared diagnostic identities, complete H1 population seals and TRAIN root view.

Fit and evaluation use the same serialized population identities. This owner preserves original labels, store order and causal prefixes; it does not admit new objective profiles or orchestrate a campaign.

1.1 Functions

Name	Description
population_identity	Reopen and seal one explicitly selected TRAIN or validation population.
validate_checkpoint_population	Verify an authenticated checkpoint’s exact admitted stage before computing.
validate_regret_validation	Require one complete 16-state chain per scene for scene-equal regret.

1.1.1 population_identity

lightning.finite_horizon.diagnostics.population.population_identity(config)

Reopen and seal one explicitly selected TRAIN or validation population.

Candidate rows remain in canonical store/chain/state order. The seal binds physical state identity, hard-valid candidate indices and native H1 labels; counts retain candidate, state, chain and scene denominators separately. Labels must cover exactly hard-valid realized support and be finite there. Unknown candidate padding is excluded rather than assigned synthetic labels.

1.1.1.1 Returns

Name	Type	Description
	dict[str, Any]	Source identity and its digest, canonical support digest, counts, and
	dict[str, Any]	the first root chain index having at least eight hard-valid labels
	dict[str, Any]	with range greater than 1e-5. The eligible index is None when
	dict[str, Any]	no root satisfies this TRAIN-canary criterion. No training is run.

1.1.1.2 Raises

Name	Type	Description
	ValueError	If the population is empty, has incomplete/nonfinite hard-valid labels, lacks support in a realized state, or repeats a physical state identity. Source readers may independently reject changed store manifests, actor carriers or descriptor contracts.

1.1.2 validate_checkpoint_population

lightning.finite_horizon.diagnostics.population.validate_checkpoint_population(
    fit_receipt,
    config,
    dataset,
    *,
    role,
)

Verify an authenticated checkpoint’s exact admitted stage before computing.

The caller authenticates receipt and checkpoint bytes through the checkpoint owner first. This pure boundary then verifies the requested configuration, reopened root bindings, scene population, descriptors and rollout digests. It does not authorize a new population or relax objective/profile admission; the DataModule and module lifecycle checks remain required before batching.

1.1.2.1 Raises

Name	Type	Description
	ValueError	If the requested stage was absent at fit time or any exact configuration or reopened population identity has drifted.

1.1.3 validate_regret_validation

lightning.finite_horizon.diagnostics.population.validate_regret_validation(
    validation,
    trainer,
)

Require one complete 16-state chain per scene for scene-equal regret.

Existing H1 logs weight states, so this diagnostic admits equal states per scene and complete hard-valid finite labels only. Truncated loaders, distributed aggregation and debug overrides are rejected prospectively. This guard establishes the meaning of the existing logged state mean; it does not change ordinary validation admission or aggregation.