1 rollouts.inspection.broad_scene
rollouts.inspection.broad_scene
Canonical tabular reductions for broad factual-rollout readouts.
The reducers keep candidate support, score availability, and policy concentration separate. They operate on rows projected by the validated rollout reader and return pandas tables for reporting; they never mutate a store or infer missing acquisitions.
1.1 Classes
| Name | Description |
|---|---|
| BroadSceneStoreFrames | Normalized factual-state and family tables from one validated store. |
1.1.1 BroadSceneStoreFrames
rollouts.inspection.broad_scene.BroadSceneStoreFrames(
states,
families,
prefixes,
marginals,
)Normalized factual-state and family tables from one validated store.
1.1.1.1 Attributes
| Name | Description |
|---|---|
| states | One row per factual decision state. |
| families | One row per factual state and persisted candidate-mixture family. |
| prefixes | Historical selected cumulative root gain at fixed factual prefixes. |
| marginals | Historical selected root-gain increment from H12 to H16. |
1.2 Functions
| Name | Description |
|---|---|
| broad_scene_store_frames | Project one validated store into reusable policy and family frames. |
| family_state_statistics | Reduce family support with explicit actor-valid and labelled gain populations. |
| historical_h12_h16_marginals | Return persisted H16-minus-H12 selected gain for factual completers. |
| historical_prefix_statistics | Project canonical factual-chain records to historical evaluator prefixes. |
1.2.1 broad_scene_store_frames
rollouts.inspection.broad_scene.broad_scene_store_frames(
reader,
*,
selection_contract,
tiny_score_iqr=0.005,
)Project one validated store into reusable policy and family frames.
Selector diagnostics reuse the producer-bound recorded-selection admission; incomplete, misaligned, or source-incompatible probability vectors remain unavailable instead of being renormalized. Candidate-family opportunity keeps actor-valid finite gains separate from training-labelled finite gains.
1.2.1.1 Parameters
| Name | Type | Description | Default |
|---|---|---|---|
| reader | RolloutZarrStoreReader | Read-only rollout-store reader. The caller owns validation so a multi-store acquisition can record every validation disposition. | required |
| selection_contract | RecordedSelectionContract | Exact producer and frozen policy contract used to validate recorded logits, probabilities, and selected identity. | required |
| tiny_score_iqr | float | Prospective small-spread threshold in root-normalized gain units. It must be finite and positive. | 0.005 |
1.2.1.2 Returns
| Name | Type | Description |
|---|---|---|
| BroadSceneStoreFrames | State and family tables preserving factual step and rollout identity. |
1.2.1.3 Raises
| Name | Type | Description |
|---|---|---|
| ValueError | If the threshold is invalid or candidate rows are missing from a rollout that has factual steps. |
1.2.2 family_state_statistics
rollouts.inspection.broad_scene.family_state_statistics(candidates)Reduce family support with explicit actor-valid and labelled gain populations.
1.2.3 historical_h12_h16_marginals
rollouts.inspection.broad_scene.historical_h12_h16_marginals(prefixes)Return persisted H16-minus-H12 selected gain for factual completers.
1.2.4 historical_prefix_statistics
rollouts.inspection.broad_scene.historical_prefix_statistics(
records,
*,
horizons=PREFIXES,
)Project canonical factual-chain records to historical evaluator prefixes.
These values reuse the old persisted root-gain evaluator. They describe prefixes of already generated factual chains and do not estimate the return of rerunning a policy with a different planning horizon.