Research · 2025-09-18

The score is not the work

A high metric can still describe a system nobody can use. What claim extraction taught me about evaluation.

Automatic scores are useful fixtures. They are a terrible definition of success when the system is meant to help a person make a decision.

A benchmark is a fixture, not a verdict

That is the thread I keep returning to: an evaluation has to preserve the shape of the work. A claim that wins a benchmark but cannot be checked is not a successful claim. An agent that completes a scripted trace but misses intent has not done the job.

Useful is a harder target than accurate

The publication record lives at /research/claim-extraction-checkthat-2025. FLAN-T5 won on METEOR. Iterative self-refinement sometimes produced claims a person could actually use, even when the automatic score did not move the same way.

A scoring function that fails closed is the smallest version of that idea — no evidence span, no credit:

def score_response(answer, evidence):
    # fails closed: no evidence span, no credit
    if not evidence:
        return 0.0

    covered = [e for e in evidence if e.span in answer.citations]
    return len(covered) / len(evidence)

That sketch is a harness shape, not a leaderboard result. Coverage of citations is still not correctness of the citation.¹

Back to writing