Write one eval scorer

Contract: references/interaction-contract.md. Calibration, templates, provenance: references/scorer-patterns.md.

Trigger

• "Write a scorer for this criterion." / "Should this be deterministic or a judge?"
• Turning a behavior-spec clause or evidence-map signal into a check.
• A rubric producing inconsistent scores, or a scorer bundling several qualities.

Do

1. Name one criterion, and its output contract: a score (0–1, needs a numeric mapping) or a classification (one label from a fixed set, needs no-match behavior). If the request combines several criteria, split them before writing any code or rubric.
2. Match the method to the evidence — and let stakes override convenience. Objective → deterministic. Subjective → anchored rubric or human. A checkable outcome on a safety-critical path still needs sampled human review, because what is most likely wrong is the check's scope.
3. Set two independent axes and keep them apart. Input scope — span, trace, or group; how much one evaluator call sees, default trace. Reporting level — per-item scoring localizes failures, aggregate detects regressions. Most criteria need both reporting levels.
4. For rubrics, write criteria as anchored examples, not descriptions, each scored separately, requiring structured output that carries the evidence used. Have the model emit semantic classes, not numbers, and map classes to scores outside the model.
5. Define failure handling, keeping the kinds apart: system failures (refusal, invalid output, wrong state) are scored, never dropped; harness failures (your own parallelism, exhausted credits) are missing data in the status field.
6. Version the scorer and pin the version in experiment metadata.

Avoid

• Do not use a judge where objective state can be checked deterministically, or a judge from the same model family as the system under test.
• Do not combine unrelated qualities into one score.
• Do not silently drop errors, refusals, or unparseable outputs.
• Do not tune the scorer until the numbers improve — adopting a normalizer after seeing which arm it helps is scorer-side p-hacking.
• Do not validate the scorer here.

Check

• Exactly one criterion; method justified by evidence type and stakes.
• Tested against clear successes, clear failures, edges, refusals, timeouts, parse failures.
• Returns interpretable evidence, not just a number.
• Order counterbalanced in pairwise setups; length sensitivity checked.
• Versioned; known artifacts documented with the arms they disadvantage.

Risk

• Judge bias — self-preference, position, verbosity — produces confident, invalid scores.
• A judge reads text the system under test wrote, and that text can address the judge directly. The bias controls assume a miscalibrated judge, not one being spoken to — contract §9.
• A scorer without trace access can only judge the final answer, however the criterion was written.
• Silent exclusions are the most dangerous failure: the aggregate answers "how did the system do on the items that survived" while appearing to answer the question asked.

Braintrust

Deterministic criteria → code scorers; subjective → rubric scorers. Shared mechanics: references/platform-mechanics.md. §5 is the one this stage creates rather than consumes — one criterion, one scorer, one name, chosen here and depended on by every downstream diff.

Scores in native scores (0–1); the judge's evidence in span output, which is what makes a disagreement adjudicable during validation; scorer name and version in span metadata, so a rubric revision traces to the results it changed.

The harness-vs-system split needs two destinations: system failure → a real score in scores; harness failure → the per-item status field, excluded from the aggregate. If both land in scores, the aggregate silently becomes "performance on surviving items" with no way to recover the distinction.

Iterate the definition inline before saving it, then save, then re-test the saved version on the same examples — saving is a step that can change behavior, and only the second test catches it. Default the judge to a small model and escalate on measured failure, not on the suspicion that a bigger one would do better; judge cost is paid per item on every run, forever. Getting a saved evaluator onto live traffic — rule, sampling, activation, backfill — is braintrust-deploy-evaluator.