1 lightning.finite_horizon.studies.q1_learning_dynamics

lightning.finite_horizon.studies.q1_learning_dynamics

Fixed-weight Q1 diagnostics with explicit state and physical-scene support.

These exploratory reductions consume already admitted H1 evaluation rows. They never fit, select checkpoints, acquire observations or infer historical gradients. Candidate errors are reduced within states before scenes receive equal weight.

1.1 Functions

Name Description
build_dynamics_tables Reduce explicit split→checkpoint roles without pooling different runs.
captured_diagnostics Collapse exact carried task diagnostics at their explicit captured update.
evaluation_frame Preserve every planned state, with null metrics for excluded states.
paired_contrasts Estimate final-minus-selected scene contrasts on one joint state mask.
read_local_wandb_history Decode one sealed local W&B transport using the installed SDK, offline.
summarize_coverage_strata Describe geometry strata, retaining native state and physical-scene units.
summarize_frame Describe admitted scene means while retaining planned and metric support.
summarize_query_coverage Census root/last query accessibility with chain and scene denominators.

1.1.1 build_dynamics_tables

lightning.finite_horizon.studies.q1_learning_dynamics.build_dynamics_tables(
    evaluations,
    *,
    tie_band,
)

Reduce explicit split→checkpoint roles without pooling different runs.

The caller admits source receipts and supplies their original rows. Split names do not confer admission. Every exclusion is visible in states/support; root mass uses original scene weights, rather than renormalizing roots first.

1.1.2 captured_diagnostics

lightning.finite_horizon.studies.q1_learning_dynamics.captured_diagnostics(
    history,
)

Collapse exact carried task diagnostics at their explicit captured update.

An update identifies a transaction, not an independent replicate. Optimizer callback fields use another cadence and are excluded. Conflicting values at one captured coordinate fail rather than being averaged; source-row bounds and transport multiplicity remain available for audit.

1.1.3 evaluation_frame

lightning.finite_horizon.studies.q1_learning_dynamics.evaluation_frame(
    rows,
    *,
    tie_band,
)

Preserve every planned state, with null metrics for excluded states.

Native FP32 residual semantics and lowest-index ties are verified by the existing state metric owner. Huber and its region fraction are descriptive functions of those same residuals, with the frozen delta one. An excluded state remains in the table and denominator; it never becomes a zero error.

1.1.4 paired_contrasts

lightning.finite_horizon.studies.q1_learning_dynamics.paired_contrasts(
    left_rows,
    right_rows,
    *,
    tie_band,
    draws=10000,
    seed=20261015,
)

Estimate final-minus-selected scene contrasts on one joint state mask.

Authoritative identities and original candidate labels must agree before excluding either side’s unsupported state. Intervals condition on these fixed weights and exposed scenes; a single scene has no population interval.

1.1.5 read_local_wandb_history

lightning.finite_horizon.studies.q1_learning_dynamics.read_local_wandb_history(
    path,
    *,
    expected_sha256,
    run_id,
)

Decode one sealed local W&B transport using the installed SDK, offline.

SDK datastore CRC validation runs before protobuf decoding. Preserve record ordinals and sparse values; never forward-fill metrics onto another update. Only scalar learning/diagnostic fields are returned, excluding private config. This is telemetry admission, not independent scientific result admission.

1.1.6 summarize_coverage_strata

lightning.finite_horizon.studies.q1_learning_dynamics.summarize_coverage_strata(
    states,
    coverage,
)

Describe geometry strata, retaining native state and physical-scene units.

Groups differ in scene and label difficulty; these means are not causal contrasts. A physical scene may contribute to both strata through different chains. Every role must cover the same declared census state identities.

1.1.7 summarize_frame

lightning.finite_horizon.studies.q1_learning_dynamics.summarize_frame(frame)

Describe admitted scene means while retaining planned and metric support.

Optional quantities, such as ordering when every label is tied, have their own scene/state denominators. State weighting is also exposed so additional correlated trajectories cannot silently change the scientific estimand.

1.1.8 summarize_query_coverage

lightning.finite_horizon.studies.q1_learning_dynamics.summarize_query_coverage(
    frame,
)

Census root/last query accessibility with chain and scene denominators.

A target-zero observation has no target-lattice point in the native sampler domain. Candidate-zero means no target or segment point is available for an action-valid row. Neither is measured-surface coverage or a learning effect.

Back to top