Results, status, and provenance

compute returns a ComputationResult. Its read-only values mapping associates each resolved feature name with a FeatureValue; execution information is separate in metadata. The snippets below use the sample from the quick start.

Mathematically undefined outputs

A feature that is undefined for an otherwise valid sample returns a FeatureValue with status invalid and an explanation rather than raising or silently producing NaN.

from orivex.result import FeatureStatus

flat = LandscapeSample(x, np.zeros(len(x)), lower=lower, upper=upper)
item = compute(flat, "ela_distr.skewness").values["ela_distr.skewness"]

item.status is FeatureStatus.INVALID  # True
item.value  # None
item.message  # 'skewness is undefined for constant objective values'

A selector that matches no registered feature is a caller error and raises instead:

from orivex.registry import UnknownFeatureSelection

compute(sample, "ela_meta.nonexistent")
# UnknownFeatureSelection: selector matched no features: ela_meta.nonexistent

Recording a computation

Record requested features, definition identifiers, preprocessing, effective options, backend, dtype, and execution settings with numerical results. Runtime belongs to metadata and is not an ELA feature.

metadata = result.metadata
print(metadata.preprocessing_fingerprint)
print(metadata.y_normalization_definition)
print(metadata.options)
print(metadata.backend, metadata.device, metadata.dtype)

for name, item in result.values.items():
    print(name, item.definition, item.status.value, item.message)

A raw sample fingerprint alone is insufficient as a result cache key. Include the preprocessing fingerprint, feature definitions, and effective execution options. See the result API for all metadata fields.