1 rollouts.inspection.sampler_rejections

rollouts.inspection.sampler_rejections

Source-bound, read-only explanations of the primary sampler rejection bucket.

These labels reconstruct the strict FP32 comparisons in MotionRealismRule; they do not decide admissibility. Unknown producers/configurations stay unresolved rather than inheriting today’s generator defaults.

1.1 Classes

Name Description
SamplerRejectionContext Original acquisition’s audited motion configuration and provenance.

1.1.1 SamplerRejectionContext

rollouts.inspection.sampler_rejections.SamplerRejectionContext(
    acquisition_manifest_sha256,
    producer_revision,
    config_sha256,
    limits,
    configured_guards,
    unavailable_reason,
)

Original acquisition’s audited motion configuration and provenance.

Explicit None limits disable a guard; absent keys are insufficient evidence. All limits and diagnostics are in metres except yaw in degrees.

1.1.1.1 Methods

Name Description
from_generation Bind the saved mixture base without applying runtime defaults.
classify Explain a primary sampler failure only when every enabled guard is known.
1.1.1.1.1 from_generation
rollouts.inspection.sampler_rejections.SamplerRejectionContext.from_generation(
    generation,
    acquisition_manifest_sha256,
)

Bind the saved mixture base without applying runtime defaults.

1.1.1.1.2 classify
rollouts.inspection.sampler_rejections.SamplerRejectionContext.classify(
    invalid_reason,
    actor_action,
    step_length_m,
    height_delta_m,
    backward_step_m,
    yaw_delta_deg,
)

Explain a primary sampler failure only when every enabled guard is known.

Persisted diagnostics and limits are cast to FP32 before strict > comparisons, matching the producer’s tensor-scalar comparison. Equality is admissible. Other active violated or unmeasured guards prevent an exclusive step/height attribution; partial flags remain available.

Back to top