Expand description
Backend-neutral model evaluation drivers.
Structs§
- Checkpoint
Parity Options - Configuration for text-checkpoint logit parity.
- Checkpoint
Parity Report - General parity report plus source artifact identities.
- Distribution
Metrics - Metrics comparing one candidate categorical distribution with a reference.
- Evaluation
Evidence - Portable evidence emitted by an evaluator, backend probe, or reference.
- Latency
Summary - Summary of repeated operation latencies.
- Logit
RowMetrics - Metrics for one vocabulary-logit row.
- Logit
Tolerance - Thresholds specialized for rows of vocabulary logits.
- Numeric
Metrics - Common metrics for floating-point vectors.
- Numeric
Tolerance - Absolute and relative thresholds for ordinary floating-point observations.
- Observation
Parity - Comparison result for one path.
- Parity
Policy - Complete policy for comparing two observation sets.
- Parity
Report - Complete backend/reference parity report.
- Parity
Rule - One selector-specific comparison rule.
- Persona
Plex Evaluation Options - Controls for one PersonaPlex comparison.
- Persona
Plex Evaluation Paths - Paths consumed and produced by one PersonaPlex comparison.
- Realtime
Trace - Completed portable outputs from one realtime request.
Enums§
- Checkpoint
Parity Error - Invalid checkpoint evidence or parity policy.
- Distribution
Error - Invalid categorical-distribution comparison.
- Evidence
Error - Invalid evaluation evidence.
- Parity
Comparison - Comparison applied to selected observation paths.
- Parity
Error - Invalid comparison policy or incompatible observations.
- Parity
Metrics - Metrics emitted for one compared observation.
- Realtime
Trace Error - Invalid portable realtime trace evidence.
Constants§
- DEFAULT_
SAMPLING_ SEED - Default deterministic sampling seed.
- DEFAULT_
TEXT_ PROMPT - Default PersonaPlex system instruction used by the evaluator.
Functions§
- compare_
checkpoint_ artifacts - Compares two checkpoint-probe report/tensor pairs using general parity.
- compare_
distributions - Compares two complete categorical logit vectors.
- compare_
observations - Compares two portable observation sets under one policy.
- encoded_
audio_ frames - Converts neutral
[batch, codebooks, frames]integer tokens to frame inputs. - observe_
f32_ tensor - Materializes a neutral tensor as F32 evaluation evidence.
- observe_
i32_ tensor - Materializes a neutral tensor as signed integer evaluation evidence.
- observe_
realtime_ frame - Converts a completed portable realtime frame into general observations.
- run_
personaplex_ quantization - Runs the complete PersonaPlex dense-versus-quantized evaluation.
- run_
realtime_ trace - Executes portable encoded frames through any realtime backend.
- summarize_
latencies - Summarizes finite, nonnegative latency samples.