1 rollouts.inspection.recipe_experiment

rollouts.inspection.recipe_experiment

Diagnostic acquisition-prefix comparisons from validated factual stores.

Prefix gains describe collection, not the value of foresight. This reducer never imputes missing endpoints, treats replicas as independent scenes, or promotes a recipe. Root camera/target identities are necessary pairing checks; prepared-cloud parity remains a separate measurement gate.

1.1 Classes

Name	Description
PrefixRecord	One factual target chain; keys exclude arm-dependent candidate tables.

1.1.1 PrefixRecord

rollouts.inspection.recipe_experiment.PrefixRecord(
    arm,
    scene,
    task_key,
    root_identity,
    gains,
    acquisition_cap,
    candidate_count=60,
    oracle_candidate_cap=60,
    termination_reason=None,
)

One factual target chain; keys exclude arm-dependent candidate tables.

1.1.1.1 Attributes

Name	Description
termination_reason	Stored stopping condition, including valid trajectories with no acquisitions.

1.2 Functions

Name	Description
paired_prefix_effects	Average matched target/replica effects within scenes before scenes.
records_from_reader	Project factual prefix records from an already admitted reader.
records_from_store	Project finite factual prefix gains only after complete store validation.

1.2.1 paired_prefix_effects

rollouts.inspection.recipe_experiment.paired_prefix_effects(
    records,
    *,
    baseline,
    challenger,
    prefix,
    candidate_budget_treatment=False,
)

Average matched target/replica effects within scenes before scenes.

prefix is an acquisition count. An incomplete chain is unavailable at later prefixes, not zero-reward evidence. Mismatched roots/caps fail rather than silently changing the comparison population. For each represented scene, average challenger-minus-baseline effects over jointly defined target/replica pairs; the conditional estimate then averages scenes equally. The exploratory 95% interval resamples those physical-scene means, using the same joint prefix support. Fewer than two paired scenes yields None. Keep the scene effects and support counts with the interval: selected completed-store populations and few scenes do not establish generalization or a promotion decision. Constant effects give a degenerate empirical interval without establishing zero population uncertainty.

1.2.2 records_from_reader

rollouts.inspection.recipe_experiment.records_from_reader(
    reader,
    *,
    arm,
    sample_key,
    replica,
    candidate_count=60,
    oracle_candidate_cap=60,
)

Project factual prefix records from an already admitted reader.

1.2.3 records_from_store

rollouts.inspection.recipe_experiment.records_from_store(
    store,
    *,
    arm,
    sample_key,
    replica,
    candidate_count=60,
    oracle_candidate_cap=60,
)

Project finite factual prefix gains only after complete store validation.

# rollouts.inspection.recipe_experiment { #aria_nbv.rollouts.inspection.recipe_experiment } `rollouts.inspection.recipe_experiment` Diagnostic acquisition-prefix comparisons from validated factual stores. Prefix gains describe collection, not the value of foresight. This reducer never imputes missing endpoints, treats replicas as independent scenes, or promotes a recipe. Root camera/target identities are necessary pairing checks; prepared-cloud parity remains a separate measurement gate. ## Classes | Name | Description | | --- | --- | | [PrefixRecord](#aria_nbv.rollouts.inspection.recipe_experiment.PrefixRecord) | One factual target chain; keys exclude arm-dependent candidate tables. | ### PrefixRecord { #aria_nbv.rollouts.inspection.recipe_experiment.PrefixRecord } ```python rollouts.inspection.recipe_experiment.PrefixRecord( arm, scene, task_key, root_identity, gains, acquisition_cap, candidate_count=60, oracle_candidate_cap=60, termination_reason=None, ) ``` One factual target chain; keys exclude arm-dependent candidate tables. #### Attributes | Name | Description | | --- | --- | | [termination_reason](#aria_nbv.rollouts.inspection.recipe_experiment.PrefixRecord.termination_reason) | Stored stopping condition, including valid trajectories with no acquisitions. | ## Functions | Name | Description | | --- | --- | | [paired_prefix_effects](#aria_nbv.rollouts.inspection.recipe_experiment.paired_prefix_effects) | Average matched target/replica effects within scenes before scenes. | | [records_from_reader](#aria_nbv.rollouts.inspection.recipe_experiment.records_from_reader) | Project factual prefix records from an already admitted reader. | | [records_from_store](#aria_nbv.rollouts.inspection.recipe_experiment.records_from_store) | Project finite factual prefix gains only after complete store validation. | ### paired_prefix_effects { #aria_nbv.rollouts.inspection.recipe_experiment.paired_prefix_effects } ```python rollouts.inspection.recipe_experiment.paired_prefix_effects( records, *, baseline, challenger, prefix, candidate_budget_treatment=False, ) ``` Average matched target/replica effects within scenes before scenes. ``prefix`` is an acquisition count. An incomplete chain is unavailable at later prefixes, not zero-reward evidence. Mismatched roots/caps fail rather than silently changing the comparison population. For each represented scene, average challenger-minus-baseline effects over jointly defined target/replica pairs; the conditional estimate then averages scenes equally. The exploratory 95% interval resamples those physical-scene means, using the same joint prefix support. Fewer than two paired scenes yields ``None``. Keep the scene effects and support counts with the interval: selected completed-store populations and few scenes do not establish generalization or a promotion decision. Constant effects give a degenerate empirical interval without establishing zero population uncertainty. ### records_from_reader { #aria_nbv.rollouts.inspection.recipe_experiment.records_from_reader } ```python rollouts.inspection.recipe_experiment.records_from_reader( reader, *, arm, sample_key, replica, candidate_count=60, oracle_candidate_cap=60, ) ``` Project factual prefix records from an already admitted reader. ### records_from_store { #aria_nbv.rollouts.inspection.recipe_experiment.records_from_store } ```python rollouts.inspection.recipe_experiment.records_from_store( store, *, arm, sample_key, replica, candidate_count=60, oracle_candidate_cap=60, ) ``` Project finite factual prefix gains only after complete store validation.