AdversaryGate gates AI coding agents with an INCONCLUSIVE verdict

A developer released AdversaryGate, a Python tool that judges coding-agent patches with three verdicts: MERGE, BLOCK or INCONCLUSIVE. It runs the baseline's tests against the new code, computes diff coverage and mutation testing, and refuses to treat unmeasured checks as passing. Since 2.8 it ships an MCP server, plus a GitHub Action, under MIT.

• Three verdicts: MERGE (0), BLOCK (1), INCONCLUSIVE (2) — unmeasured claims never clear a floor
• Tests come from the baseline tree, so a patch that rewrites assertions cannot pass
• Mutation score uses the 80% Wilson interval lower bound, not the raw ratio
• Since 2.8 an MCP server lets agents call the gate; policy and base ref are set server-side