Judge calibration report
Judge: deterministic-lexical (config ff1ad7874e00)
- Probes: 66
- Raw agreement with human labels: 0.955 (threshold 0.8)
- Cohen's Îș: 0.906 (threshold 0.6) â â meets
| Operation | Agree | Agreement |
|---|---|---|
| contains | 22/22 | 1.00 |
| entails | 27/27 | 1.00 |
| equivalent | 14/17 | 0.82 |
Disagreements: e2-es-yellow-leaves-synonym-paraphrase, e16-es-brown-tips-synonym-paraphrase, e17-en-root-rot-synonym-paraphrase
The deterministic lexical judge is the reproducible offline floor: lexical coverage plus a negation/antonym polarity guard, not a general-purpose semantic judge. It still misses morphological synonyms and paraphrase that share little surface vocabulary, which is why production gates should ultimately be backed by the calibrated LLM judge (
--judge llm) as the probe set grows. Pass--gatetosprout calibrateto fail the build below threshold; run without it to report only.