File size: 1,663 Bytes
d4558b2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 | # Evaluation protocol
Evaluate translation as grounded data transformation, not sentence similarity.
## Required test pools
- frozen bilingual gold rows for every supported language and construction;
- video-grounded rows stratified by action, object, hand configuration,
environment, state transition, confidence, and source;
- contrast pairs for open/close, source/destination, left/right, success/error,
observation/directive, singular/plural, negation, and temporal order;
- malformed and adversarial records; and
- rare vocabulary and low-confidence strata that cannot be hidden by frequency.
## Metrics
Report separately by language and stratum:
- supported-claim precision and recall;
- critical, major, and minor errors per 1,000 rows;
- action, object, hand, state, temporal, and provenance accuracy;
- bilingual adequacy and naturalness;
- withheld/abstention rate;
- multilingual video retrieval and temporal-localization performance; and
- constant-answer, language-blind, and source-only baselines.
## Recommended release gates
- at least 99% of expressed claims supported by evidence;
- fewer than 0.5% critical semantic errors;
- at least 95% canonical-claim equivalence across language surfaces;
- mean bilingual adequacy at least 4.3/5 and naturalness at least 4.0/5;
- no failed action, source, factory, or language-hazard stratum hidden by the
aggregate; and
- multilingual retrieval within 5% relative of the English condition on the
same frozen relevance judgments.
These are proposed acceptance thresholds, not scores achieved by this release.
The controlled conformance matrix is too small to establish production quality.
|