1 lightning.finite_horizon.studies.first_model_evidence

lightning.finite_horizon.studies.first_model_evidence

Render the sealed joint R0/crop-R1 commissioning study without running fits.

Scene dots show descriptive variation; intervals resample physical scenes with shared draws across the two recipes. Selected and final checkpoints remain separate, and partial histories never become completed evaluation evidence.

1.1 Classes

Name Description
FirstModelCellDisposition Completion evidence for a registered cell, independent of its telemetry.
FirstModelLearningObservation One retained loss observation; updates are not independent scene samples.
FirstModelSceneContrast One recipe/checkpoint comparison reduced over paired physical scenes.
FirstModelStudy Admitted commissioning evidence; scientific summaries are serialized snapshots.

1.1.1 FirstModelCellDisposition

lightning.finite_horizon.studies.first_model_evidence.FirstModelCellDisposition(
    cell,
    disposition,
    seed,
)

Completion evidence for a registered cell, independent of its telemetry.

1.1.2 FirstModelLearningObservation

lightning.finite_horizon.studies.first_model_evidence.FirstModelLearningObservation(
    cell,
    seed,
    optimizer_update,
    event,
    loss,
)

One retained loss observation; updates are not independent scene samples.

1.1.3 FirstModelSceneContrast

lightning.finite_horizon.studies.first_model_evidence.FirstModelSceneContrast(
    cell,
    checkpoint_role,
    metric,
    recipe,
    model_mean,
    constant_mean,
)

One recipe/checkpoint comparison reduced over paired physical scenes.

1.1.4 FirstModelStudy

lightning.finite_horizon.studies.first_model_evidence.FirstModelStudy(
    study,
    snapshot_json,
    _histories,
    _native_inputs_json=None,
)

Admitted commissioning evidence; scientific summaries are serialized snapshots.

1.1.4.1 Attributes

Name Description
provenance Expose admitted producer identities without leaking mutable protocol nodes.
cell_dispositions Keep unavailable completion evidence distinct from a completed fit.
scene_contrasts Expose the standalone table’s exact scene-equal means at its recipe grain.
learning_observations Return immutable train/validation loss points from admitted local history.
native_inputs Return detached native rows; aggregate summaries cannot replace them.
histories Detach nested identities/events so renderer mutations cannot alter evidence.
snapshot Return a detached snapshot so renderers cannot alter admitted evidence.

1.2 Functions

Name Description
acquire_first_model_inputs Acquire native Q1 receipts once, preserving exact hashes and schema tokens.
admit_archived_first_model Admit original Q1 receipts from an externally authenticated portable archive.
admit_first_model Admit frozen Q1 support, completed receipts and partial-history dispositions.
baseline_summary Estimate absolute baseline performance, resampling paired physical scenes.
cloud_publication Distinguish local evaluation from a recorded, result-bound cloud readback.
completed_cell Admit completed selected/final receipts; never display stale failed-fit rows.
constant_loss_reference Reconstruct native delta-one Huber with equal-state, recipe and scene means.
digest Hash canonical JSON using the scientific producer’s encoding.
history_for Use the canonical partial-history reader and frozen dataset/model bindings.
identity Preserve scene, ordered store, rollout and realized state identities.
paired_rows Reject wrong support before joining predictions by authoritative identity.
prefix_contrasts Separate root and later-prefix diagnostics without changing full-chain gates.
read Read an explicit receipt; missing evidence is handled by the caller.
scene_contrasts Estimate equal-scene/equal-recipe contrasts with physical-scene bootstrap.
sha Bind a rendered measurement to its exact retained receipt bytes.
validate_first_model_schemas Reject missing/unknown protocol, cell and evaluation formats before reduction.
validate_loss_control Admit only the frozen pure H1 regression-Huber objective for loss overlays.
validated_constant Bind ordered recipes, source provenance and original labels to admission.

1.2.1 acquire_first_model_inputs

lightning.finite_horizon.studies.first_model_evidence.acquire_first_model_inputs(
    study,
    *,
    completed_cells=None,
)

Acquire native Q1 receipts once, preserving exact hashes and schema tokens.

Completion selection is caller-owned: standalone post-hoc admits only its full-seed41 cell, while readiness admits all registered cells. Schema checks retain their original campaign-wide coverage without admitting other fits.

1.2.2 admit_archived_first_model

lightning.finite_horizon.studies.first_model_evidence.admit_archived_first_model(
    archive,
    prefix,
)

Admit original Q1 receipts from an externally authenticated portable archive.

1.2.3 admit_first_model

lightning.finite_horizon.studies.first_model_evidence.admit_first_model(study)

Admit frozen Q1 support, completed receipts and partial-history dispositions.

1.2.4 baseline_summary

lightning.finite_horizon.studies.first_model_evidence.baseline_summary(
    constant,
    recipes,
    *,
    expected_scenes,
)

Estimate absolute baseline performance, resampling paired physical scenes.

Reuse the existing complete recipe-by-scene admission and reductions. The baseline interval resamples absolute scene means, never zero self-contrasts. Macro observations average R0 and crop R1 within each scene before sampling.

1.2.5 cloud_publication

lightning.finite_horizon.studies.first_model_evidence.cloud_publication(
    study,
    cell,
    *,
    local_completed,
    _inputs=None,
)

Distinguish local evaluation from a recorded, result-bound cloud readback.

This renderer performs no network request. Verified means the retained publication receipt records byte readback of the same local result, not that hosted state was checked again while rendering.

1.2.6 completed_cell

lightning.finite_horizon.studies.first_model_evidence.completed_cell(
    study,
    cell,
    protocol,
    constant,
    recipes,
    *,
    _inputs=None,
)

Admit completed selected/final receipts; never display stale failed-fit rows.

1.2.7 constant_loss_reference

lightning.finite_horizon.studies.first_model_evidence.constant_loss_reference(
    constant,
    recipes,
    *,
    expected_scenes,
)

Reconstruct native delta-one Huber with equal-state, recipe and scene means.

Labels and predictions are converted to float32 before candidate losses and within-state means, matching native regression evaluation. Dependent states are reduced before float64 descriptive scene aggregation and bootstrap. The interval belongs to the constant baseline, not the model loss curve.

1.2.8 digest

lightning.finite_horizon.studies.first_model_evidence.digest(value)

Hash canonical JSON using the scientific producer’s encoding.

1.2.9 history_for

lightning.finite_horizon.studies.first_model_evidence.history_for(
    study,
    cell,
    protocol,
)

Use the canonical partial-history reader and frozen dataset/model bindings.

1.2.10 identity

lightning.finite_horizon.studies.first_model_evidence.identity(row)

Preserve scene, ordered store, rollout and realized state identities.

1.2.11 paired_rows

lightning.finite_horizon.studies.first_model_evidence.paired_rows(
    learned,
    constant,
)

Reject wrong support before joining predictions by authoritative identity.

Equal counts do not establish pairing. Candidate indices and original labels must agree within each state, including their ordering, before subtraction.

1.2.12 prefix_contrasts

lightning.finite_horizon.studies.first_model_evidence.prefix_contrasts(
    learned,
    constant,
    recipes,
    *,
    expected_scenes,
)

Separate root and later-prefix diagnostics without changing full-chain gates.

Preserve original state indices and candidate pairing. Each later-prefix observation averages its fifteen dependent states within recipe, then the two recipes within physical scene. Root observations use state zero only. Both phases use the existing shared physical-scene bootstrap seed/draws.

1.2.13 read

lightning.finite_horizon.studies.first_model_evidence.read(path)

Read an explicit receipt; missing evidence is handled by the caller.

1.2.14 scene_contrasts

lightning.finite_horizon.studies.first_model_evidence.scene_contrasts(
    learned,
    constant,
    store_recipes,
    *,
    expected_scenes,
    expected_states=16,
    expected_state_indices=None,
)

Estimate equal-scene/equal-recipe contrasts with physical-scene bootstrap.

Reduce dependent states within each recipe and scene first. The same 10,000 draws of scenes are used for R0, R1 and their equal-recipe macro average. These are exploratory development intervals, not seed replication or an untouched-test claim. Differences are learned minus TRAIN-fitted constant.

1.2.15 sha

lightning.finite_horizon.studies.first_model_evidence.sha(path)

Bind a rendered measurement to its exact retained receipt bytes.

1.2.16 validate_first_model_schemas

lightning.finite_horizon.studies.first_model_evidence.validate_first_model_schemas(
    study,
)

Reject missing/unknown protocol, cell and evaluation formats before reduction.

Schema tokens are scientific admission, independent of self-consistent hashes. Low-level historical helpers remain available for focused support diagnostics.

1.2.17 validate_loss_control

lightning.finite_horizon.studies.first_model_evidence.validate_loss_control(
    cells,
)

Admit only the frozen pure H1 regression-Huber objective for loss overlays.

1.2.18 validated_constant

lightning.finite_horizon.studies.first_model_evidence.validated_constant(
    study,
    protocol,
    *,
    _inputs=None,
)

Bind ordered recipes, source provenance and original labels to admission.

Back to top