rc-bench

Purpose

A rawcode extension that turns "is this prompt change better?" into a number. It runs the repo's bench/ — a paired HumanEval pass@1 A/B graded by executable unit tests (no LLM judge) — and reports correctness and token deltas as separate axes. Use it when a persona/prompt/config change needs evidence, not an opinion.

When to use

• Before claiming a system-prompt or persona change "helps".
• As a quantitative step inside sdd-verify for a change that alters agent behavior.
• To sanity-check that a conciseness tweak did not lower correctness (it can — see the Anthropic conciseness postmortem).

How to run

bench/run.sh 40        # 40 HumanEval problems, baseline vs the rawcode prompt

It downloads HumanEval on first run, generates one greedy completion per problem per arm via claude -p, runs each problem's own tests, then prints:

• Correctness (pass@1) per arm, the paired difference, a 95% CI, and a McNemar exact p-value. p > 0.05 = no significant change.
• Output tokens per arm and the paired reduction with a bootstrap 95% CI.

Reading the result honestly

• A win = tokens down with pass@1 not significantly down. Report both.
• Never blend the two into one score — length and correctness are different axes.
• Watch for grading artifacts: an aggressive persona can make the model return a terse confirmation ("Done") or act agentically instead of printing code, which a naive extractor miscounts as a failure. If pass@1 drops, inspect the raw outputs before believing it — the drop may be format, not quality.
• Small N (40, single sample) → wide correctness CI. Raise N or samples for pass@k.

Compose with Gentle AI

• Inside sdd-verify: cite the bench/ numbers (with CI) as the quantitative evidence, alongside the spec-conformance check.
• Persist the run summary under the active artifact store's verify-report key — no new backend.