1 vin.types.candidate_features

vin.types.candidate_features

Typed candidate-readout values, support masks, and static feature specifications.

1.1 Classes

Name	Description
QhCandidateFeatureSpec	Static feature order and units, available without constructing a model.
QhCandidateFeatures	Candidate-aligned pooled EVL evidence with explicit missingness.

1.1.1 QhCandidateFeatureSpec

vin.types.candidate_features.QhCandidateFeatureSpec(names, units)

Static feature order and units, available without constructing a model.

Feature names align exactly with the last axis of QhCandidateFeatures. Readout producers order blocks by region, then channel, then statistic. Region labels distinguish globally pooled root evidence from target-lattice and candidate-ray reductions; they do not by themselves imply that every feature is candidate-conditioned.

1.1.1.1 Attributes

Name	Description
names	Unique ordered `region.channel.statistic` names, one per feature.
units	Physical-unit labels aligned one-to-one with `names`.
output_dim	Return the feature width without runtime tensor execution.

1.1.2 QhCandidateFeatures

vin.types.candidate_features.QhCandidateFeatures(values, available, spec)

Candidate-aligned pooled EVL evidence with explicit missingness.

Instances may contain globally pooled summary features, spatially pooled target-lattice and candidate-ray features, or another representation named by spec. Summary values depend on root EVL evidence and are broadcast over states and candidates. Spatial target blocks additionally depend on target geometry but are shared by candidates; ray blocks also depend on the corresponding candidate pose. None of these representations encode authoritative action validity.

The B,S,N,F layout means batch, realized/padded chain state, stored candidate row, then feature. Missing source evidence, unsupported spatial samples, and padded rows are zero-filled. An observed zero remains distinct because its availability is true. Consumers must carry available with values and interpret columns through the exact spec.

1.1.2.1 Attributes

Name	Description
values	`Tensor["B S N F", float32]` finite pooled values in `spec` order.
available	`Tensor["B S N F", bool]` measurement support, distinct from action validity.
spec	Static feature order and units; `spec.output_dim` equals `F`.

# vin.types.candidate_features { #aria_nbv.vin.types.candidate_features } `vin.types.candidate_features` Typed candidate-readout values, support masks, and static feature specifications. ## Classes | Name | Description | | --- | --- | | [QhCandidateFeatureSpec](#aria_nbv.vin.types.candidate_features.QhCandidateFeatureSpec) | Static feature order and units, available without constructing a model. | | [QhCandidateFeatures](#aria_nbv.vin.types.candidate_features.QhCandidateFeatures) | Candidate-aligned pooled EVL evidence with explicit missingness. | ### QhCandidateFeatureSpec { #aria_nbv.vin.types.candidate_features.QhCandidateFeatureSpec } ```python vin.types.candidate_features.QhCandidateFeatureSpec(names, units) ``` Static feature order and units, available without constructing a model. Feature names align exactly with the last axis of :class:`QhCandidateFeatures`. Readout producers order blocks by region, then channel, then statistic. Region labels distinguish globally pooled root evidence from target-lattice and candidate-ray reductions; they do not by themselves imply that every feature is candidate-conditioned. #### Attributes | Name | Description | | --- | --- | | [names](#aria_nbv.vin.types.candidate_features.QhCandidateFeatureSpec.names) | Unique ordered ``region.channel.statistic`` names, one per feature. | | [units](#aria_nbv.vin.types.candidate_features.QhCandidateFeatureSpec.units) | Physical-unit labels aligned one-to-one with ``names``. | | [output_dim](#aria_nbv.vin.types.candidate_features.QhCandidateFeatureSpec.output_dim) | Return the feature width without runtime tensor execution. | ### QhCandidateFeatures { #aria_nbv.vin.types.candidate_features.QhCandidateFeatures } ```python vin.types.candidate_features.QhCandidateFeatures(values, available, spec) ``` Candidate-aligned pooled EVL evidence with explicit missingness. Instances may contain globally pooled summary features, spatially pooled target-lattice and candidate-ray features, or another representation named by ``spec``. Summary values depend on root EVL evidence and are broadcast over states and candidates. Spatial target blocks additionally depend on target geometry but are shared by candidates; ray blocks also depend on the corresponding candidate pose. None of these representations encode authoritative action validity. The ``B,S,N,F`` layout means batch, realized/padded chain state, stored candidate row, then feature. Missing source evidence, unsupported spatial samples, and padded rows are zero-filled. An observed zero remains distinct because its availability is true. Consumers must carry ``available`` with ``values`` and interpret columns through the exact ``spec``. #### Attributes | Name | Description | | --- | --- | | [values](#aria_nbv.vin.types.candidate_features.QhCandidateFeatures.values) | ``Tensor["B S N F", float32]`` finite pooled values in ``spec`` order. | | [available](#aria_nbv.vin.types.candidate_features.QhCandidateFeatures.available) | ``Tensor["B S N F", bool]`` measurement support, distinct from action validity. | | [spec](#aria_nbv.vin.types.candidate_features.QhCandidateFeatures.spec) | Static feature order and units; ``spec.output_dim`` equals ``F``. |