proofbench

A config-driven eval harness that grades and self-improves headless agent skills against ground-truth corpora. You plant tasks you already have the answers to, run an agent skill across them, and grade what it changed against the known answer with an LLM judge. Then you let the harness rewrite the skill from its own misses and keep the rewrite only if the score holds.

It started as the harness behind crg-debug, a graph-driven debugging skill. The methodology generalizes to any skill whose output you can compare to a known answer.

Why it exists

A skill is only as trustworthy as the proof that it works. Anyone can write a prompt that sounds like a methodology. The honest way to know is to measure it against ground truth, on the model weak enough to embarrass it. Two ideas carry the whole design, both learned the hard way (see METHODOLOGY.md):

1. Fail loud or do not measure. Every empty or non-numeric grade is a hard stop. An eval that fails open does not just miss data, it manufactures false confidence.
2. The weak model is the signal, not the noise. A self-improving loop learns only from misses. Frontier models on easy tasks miss nothing, so they teach nothing. The weak leg is the curriculum.

Install

uv tool install proofbench      # or: uvx proofbench

Requires the claude CLI on PATH for the default runner and for the judge.

Quickstart (demo)

git clone https://github.com/CodeBlackwell/proofbench && cd proofbench
bash examples/sample-corpus/build.sh        # builds two toy repos with buggy + fixed branches
uvx --from . proofbench run --demo          # eval an agent over them, graded vs the answers

How it works

One YAML config declares everything domain-specific; the engine is generic. The mode is implied by what the config contains:

• a corpus + runner + judge gives you a bench (run)
• adding a subject + synth unlocks the self-improving loop (optimize)

subject: ~/.claude/skills/crg-debug/SKILL.md   # optional; omit for pure-eval mode
runner: claude                                  # default adapter; any executable works
models: [opus, sonnet, haiku]                   # the driver sweep; the weak leg is the signal
judge_model: opus                               # held constant; never let a model grade itself
objective: macro_recall                         # the metric the keep/revert gate reads
judge: prompts/judge.md
synth: prompts/synth.md                          # present => `optimize` is available
corpus:
  - name: primes
    path: examples/sample-corpus/repos/primes
    invoke: "Find and fix the bug in this repository."
    default_branch: buggy
    answer_branch: fixed

proofbench run      --config proofbench.yaml    # eval + scoreboard
proofbench optimize --config proofbench.yaml    # baseline -> synth -> re-run -> keep|revert

The keep/revert gate

optimize runs a baseline on the weak leg, asks an LLM to rewrite the subject from the graded misses, re-runs, and keeps the rewrite only if the objective did not regress. The decision is the harness comparing two numbers, never the model's own claim of success.

Adapters

• Runner (runner:): claude is built in. Any other value is an executable invoked as <runner> <invoke> <model>, so you can drive aider, codex, or a custom agent.
• Capture: the git-diff default resets a repo to its default branch and snapshots what the agent changed, excluding dependency and cache trees so they never pollute or balloon the judge prompt.

Bring your own corpus

The bundled corpus is two MIT toy repos for the demo. Point corpus: at your own repos (each with a broken default branch and a fixed solution branch). See CORPUS.md for how to curate tasks that actually discriminate and do not leak their answers to the judge.

License

MIT.