# Evaluation protocol (draft)

## Run conditions

- Pin the exact model repository and commit hash before inference.
- Use the same chat template and decoding settings for every item; record them in the run manifest.
- For deterministic comparisons, use greedy decoding (do_sample=false, temperature=0). If the deployed model requires sampling, publish the seed and repeat each item three times.
- Keep prompts, raw outputs, model revision, runtime, and timestamps in a separate results file. Do not put outputs into the gold dataset.
- Do not use examples from the model's SFT datasets as hidden evaluation items.

## Scoring

### Identity (100)

Score each item 0–2: 2 = factually grounded and calibrated; 1 = partly correct or overly vague; 0 = false identity, accepts a false premise, or fabricates an unsupported value. Report the mean and rates of false-premise correction and unsupported-claim fabrication separately.

### French ↔ Baoulé translation (600)

Use two independent fluent Baoulé reviewers for each gold translation. Record dialect/script preferences and accepted variants. Score adequacy, meaning preservation, omissions/additions, and fluency on a 0–4 human rubric. Report direction-level results separately. chrF++ may be reported as a secondary diagnostic after the references are adjudicated; do not treat it as the only score.

### General knowledge (300)

Use exact multiple-choice accuracy from the imported gold answer. Report accuracy by subject and the aggregate. Keep the original MMLU benchmark's result separate from this fixed 300-item sample.

## Translation review rubric (0–4)

- 4: meaning preserved, natural Baoulé/French, no material error.
- 3: meaning preserved with minor wording or spelling issues.
- 2: partly correct; one material omission, addition, or mistranslation.
- 1: mostly incorrect or difficult to understand.
- 0: unrelated, empty, or wrong-language response.

## Run report

Publish total and per-track scores, item counts, reviewer agreement, model revision, decoding settings, and a short error analysis. Do not publish a single combined score without showing the component scores.