Specialist-model research desk

Direct questions

Which metrics should a buyer insist on?

Frontier Development research deskSection 01 · Page 6 of 14Last updated 26 October 2026

Define success at the business-task level, with explicit failure constraints. Averages alone can hide rare but costly errors.

Workflow Primary acceptance measure Essential companion measures
Invoice extraction Entire document satisfies agreed field and business rules Critical-field errors, abstention, review minutes, duplicate handling
Support classification Correct destination and permitted next action Rare-intent recall, wrong escalation, tenant-policy errors
Support drafting Reviewer accepts answer under the rubric Unsupported statements, edits required, review time
Autonomous resolution Correct permitted outcome completed Unauthorised actions, reversals, unresolved cases, escalation quality

Show denominators, sample composition and uncertainty. For correlated records, estimate uncertainty at a suitable cluster level; ask an evaluation specialist to check the design if the decision turns on a small difference. Anthropic’s evaluation guidance distinguishes code, model and human graders. The desk’s recommendation is to use deterministic checks where possible and calibrate subjective grading against adjudicated examples.

TRACE THE EVIDENCE

Sources cited on this page

  1. Anthropic’s evaluation guidance →www.anthropic.com