1 lightning.finite_horizon.studies.first_model_analysis

lightning.finite_horizon.studies.first_model_analysis

Explore sealed Q1 predictions without fitting or changing checkpoint selection.

Run python -m scripts.first_model_posthoc --study ... --output .... States and recipes are dependent within a physical scene. All intervals hold weights fixed and are exploratory; no finding changes the frozen readiness gate.

1.1 Classes

Name Description
FirstModelAnalysis Fixed-weight, scene-paired post-hoc product with detached state observations.
FirstModelPhysicalSceneContrast Fixed-weight paired scene observations and their existing bootstrap summary.

1.1.1 FirstModelAnalysis

lightning.finite_horizon.studies.first_model_analysis.FirstModelAnalysis(
    summary_json,
    records_json,
)

Fixed-weight, scene-paired post-hoc product with detached state observations.

1.1.1.1 Attributes

Name Description
physical_scene_contrasts Project existing reductions without refitting or resampling observations.

1.1.2 FirstModelPhysicalSceneContrast

lightning.finite_horizon.studies.first_model_analysis.FirstModelPhysicalSceneContrast(
    metric,
    comparison,
    seed,
    scene_ids,
    scene_values,
    mean,
    ci95,
    scene_count,
)

Fixed-weight paired scene observations and their existing bootstrap summary.

1.2 Functions

Name Description
admit_archived_first_model_analysis Compute fixed-weight post-hoc evidence using only authenticated portable inputs.
admit_first_model_analysis Validate original support before computing the fixed-weight analysis.
analyze Bind sealed receipts before deriving state tables and paired comparisons.
estimate Describe equal-scene values and conditional percentile mean uncertainty.
scene_matrix Reduce states then recipes; reject missing recipe/scene/role coverage.
state_metrics Describe original supported candidates, preserving global action indices.
trajectory Explore prefix regret contrasts with pointwise and simultaneous bands.

1.2.1 admit_archived_first_model_analysis

lightning.finite_horizon.studies.first_model_analysis.admit_archived_first_model_analysis(
    archive,
    prefix,
)

Compute fixed-weight post-hoc evidence using only authenticated portable inputs.

1.2.2 admit_first_model_analysis

lightning.finite_horizon.studies.first_model_analysis.admit_first_model_analysis(
    study,
)

Validate original support before computing the fixed-weight analysis.

1.2.3 analyze

lightning.finite_horizon.studies.first_model_analysis.analyze(study)

Bind sealed receipts before deriving state tables and paired comparisons.

1.2.4 estimate

lightning.finite_horizon.studies.first_model_analysis.estimate(values, draws)

Describe equal-scene values and conditional percentile mean uncertainty.

1.2.5 scene_matrix

lightning.finite_horizon.studies.first_model_analysis.scene_matrix(
    frame,
    metric,
    scenes,
)

Reduce states then recipes; reject missing recipe/scene/role coverage.

Return a float64 array with axes (role, scene). There must be exactly two recipes per scene for every role; an absent observation is never zero.

1.2.6 state_metrics

lightning.finite_horizon.studies.first_model_analysis.state_metrics(
    row,
    tie_band,
)

Describe original supported candidates, preserving global action indices.

Uniform regret is the analytic expected regret of selecting each admitted action with equal probability. It needs no prediction or fitted parameter. The fraction of practically better actions uses the original utility band; ties do not count against a choice. Calibration spread uses population SD. Candidate-mean signed error is a common offset: MSE equals its square plus centered error MSE. This descriptive decomposition uses the same supported FP32 residuals and does not fit a held-out calibration. Single-candidate margins are missing; top-three overlap uses the available candidate count and deterministic global-index tie breaking.

1.2.7 trajectory

lightning.finite_horizon.studies.first_model_analysis.trajectory(
    frame,
    scenes,
    draws,
)

Explore prefix regret contrasts with pointwise and simultaneous bands.

Shared physical-scene draws cover the 32 selected/final × 16-prefix macro contrasts. The simultaneous band uses the bootstrap maximum absolute standardized centered mean, with fixed empirical scene standard errors. It is approximate and conditional, not a post-selection confirmatory test.

Back to top