Ante

Don't ask who wrote it. Ask what it proves.

Ante asks a pull request one question: is the claim it makes independently reproducible? It never asks whether a person or a model wrote the patch — that question cannot be answered reliably, and answering it badly punishes people who have done nothing wrong.

GitHub Benchmark dataset Claim-alignment judge

starting Python in your browser…

Load a case from the benchmark The claim (issue / PR body) File path File — before File — after Attached test (the evidence)

Result

Load a benchmark case on the left, or paste your own change, then press Check the evidence.

This page runs entirely in your browser. Nothing you paste is uploaded anywhere. That also means there is no process here to run a test in, so this page performs the static half of the protocol: what the claim asserts, and whether the attached test is a program that could fail for the claimed reason. It is the half that catches fabricated evidence.

The other half — running your test against both commits, tracing which changed lines it reaches, re-running the project's suite, checking a benchmark's confidence interval — needs a machine:

pip install git+https://github.com/NagaYu/ante
ante review --base main --head HEAD --claim pr-body.md

or add - uses: NagaYu/ante/action@v1 to your CI, where it runs in about half a second per pull request and does not fail your build by default.

What Ante refuses to build: no authorship or “AI-written” detection, no contributor scores or rankings, no public directory of anyone's history, and no way to record a negative about a person. Every pull request in the benchmark is synthetic; no real contributor or repository is evaluated.