1 rollouts.inspection.candidate_population

rollouts.inspection.candidate_population

Scene-balanced descriptive reductions of store-qualified candidate evidence.

The input is the complete admitted scalar audit population, never its display sample. Original reasons and masks remain untouched. Equal scene weight applies inside each supported acquisition/family/population; occurrence counts are a separate estimand. No store reads, admission decisions or plotting occur here.

1.1 Attributes

Name	Description
UNAVAILABLE_BINDING	Display-only stratum for missing bindings; source nulls remain unchanged.

1.2 Classes

Name	Description
CandidatePopulationDiagnostics	Detached scalar population and its precomputed descriptive tables.

1.2.1 CandidatePopulationDiagnostics

rollouts.inspection.candidate_population.CandidatePopulationDiagnostics(
    rows,
    primary,
    triggered,
    survival,
    distance,
    clearance,
    metric_support,
    receipt,
    families,
    acquisitions,
)

Detached scalar population and its precomputed descriptive tables.

Frames are owned by this result (not aliases of input frames). A frozen dataclass prevents field replacement, not pandas mutation; consumers must treat frames as read-only. primary includes original/refined variants; distance and clearance include the full population plus each family. Support columns distinguish finite observations from missing/nonfinite data. At each materialized acquisition, available_candidates counts candidates before the population mask; materialized counts rows after that mask. A zero selected count is distinct from selected rows with missing distances.

1.2.1.1 Attributes

Name	Description
sampler_attribution	Return occurrence counts by original source guards and explanation.

1.3 Functions

Name	Description
filter_candidate_population	Select exact scalar strata and recompute without acquiring source stores.
population_candidate_diagnostics	Reduce complete source-qualified audit rows without changing their labels.

1.3.1 filter_candidate_population

rollouts.inspection.candidate_population.filter_candidate_population(
    evidence,
    *,
    cohort=None,
    contract=None,
    family=None,
)

Select exact scalar strata and recompute without acquiring source stores.

An unspecified binding retains that axis; therefore mixed remaining axes continue to require partition rather than being silently pooled.

1.3.2 population_candidate_diagnostics

rollouts.inspection.candidate_population.population_candidate_diagnostics(rows)

Reduce complete source-qualified audit rows without changing their labels.

Required identity is store_path/store_id, rollout_row_id, step_row_id, candidate_row_id and scene. Repeated identities fail rather than silently duplicating support. A mixed known cohort or compatibility contract fails; absent identities remain explicitly unavailable in the receipt. Mixed populations return scalar rows and a partition-required receipt without pooled tables.

Scene-balanced distances assign each finite row weight 1/n_scene within acquisition/population/family. Quantiles interpolate midpoint cumulative weights. Clearance uses the corresponding exact weighted ECDF, retaining all finite tails and measured zero. Neither is a confidence interval.

# rollouts.inspection.candidate_population { #aria_nbv.rollouts.inspection.candidate_population } `rollouts.inspection.candidate_population` Scene-balanced descriptive reductions of store-qualified candidate evidence. The input is the complete admitted scalar audit population, never its display sample. Original reasons and masks remain untouched. Equal scene weight applies inside each supported acquisition/family/population; occurrence counts are a separate estimand. No store reads, admission decisions or plotting occur here. ## Attributes | Name | Description | | --- | --- | | [UNAVAILABLE_BINDING](#aria_nbv.rollouts.inspection.candidate_population.UNAVAILABLE_BINDING) | Display-only stratum for missing bindings; source nulls remain unchanged. | ## Classes | Name | Description | | --- | --- | | [CandidatePopulationDiagnostics](#aria_nbv.rollouts.inspection.candidate_population.CandidatePopulationDiagnostics) | Detached scalar population and its precomputed descriptive tables. | ### CandidatePopulationDiagnostics { #aria_nbv.rollouts.inspection.candidate_population.CandidatePopulationDiagnostics } ```python rollouts.inspection.candidate_population.CandidatePopulationDiagnostics( rows, primary, triggered, survival, distance, clearance, metric_support, receipt, families, acquisitions, ) ``` Detached scalar population and its precomputed descriptive tables. Frames are owned by this result (not aliases of input frames). A frozen dataclass prevents field replacement, not pandas mutation; consumers must treat frames as read-only. ``primary`` includes original/refined variants; ``distance`` and ``clearance`` include the full population plus each family. Support columns distinguish finite observations from missing/nonfinite data. At each materialized acquisition, ``available_candidates`` counts candidates before the population mask; ``materialized`` counts rows after that mask. A zero selected count is distinct from selected rows with missing distances. #### Attributes | Name | Description | | --- | --- | | [sampler_attribution](#aria_nbv.rollouts.inspection.candidate_population.CandidatePopulationDiagnostics.sampler_attribution) | Return occurrence counts by original source guards and explanation. | ## Functions | Name | Description | | --- | --- | | [filter_candidate_population](#aria_nbv.rollouts.inspection.candidate_population.filter_candidate_population) | Select exact scalar strata and recompute without acquiring source stores. | | [population_candidate_diagnostics](#aria_nbv.rollouts.inspection.candidate_population.population_candidate_diagnostics) | Reduce complete source-qualified audit rows without changing their labels. | ### filter_candidate_population { #aria_nbv.rollouts.inspection.candidate_population.filter_candidate_population } ```python rollouts.inspection.candidate_population.filter_candidate_population( evidence, *, cohort=None, contract=None, family=None, ) ``` Select exact scalar strata and recompute without acquiring source stores. An unspecified binding retains that axis; therefore mixed remaining axes continue to require partition rather than being silently pooled. ### population_candidate_diagnostics { #aria_nbv.rollouts.inspection.candidate_population.population_candidate_diagnostics } ```python rollouts.inspection.candidate_population.population_candidate_diagnostics(rows) ``` Reduce complete source-qualified audit rows without changing their labels. Required identity is store_path/store_id, rollout_row_id, step_row_id, candidate_row_id and scene. Repeated identities fail rather than silently duplicating support. A mixed known cohort or compatibility contract fails; absent identities remain explicitly unavailable in the receipt. Mixed populations return scalar rows and a partition-required receipt without pooled tables. Scene-balanced distances assign each finite row weight ``1/n_scene`` within acquisition/population/family. Quantiles interpolate midpoint cumulative weights. Clearance uses the corresponding exact weighted ECDF, retaining all finite tails and measured zero. Neither is a confidence interval.