Tamperward

Appearance

TamperwardThe deterministic agent-integrity gate

AI coding agents optimize for "the command succeeded," not "the change is trustworthy." Tamperward blocks the class of shortcut they take to force checks green — measured, not asserted.

Get started
Read the launch post
GitHub
TamperwardTamperward

  • Deterministic, no model calls

    A ruleset evaluated on the actual diff and commands. There is no LLM judge to persuade — a gate an agent can argue with is not a gate.

  • Enforced at every stage, CI the authority

    Inside the agent's loop (a Claude Code PreToolUse deny plus an end-of-turn sweep, holding under bypassPermissions), at pre-commit, and in CI — the authority, reading the policy from the merge-base so a PR cannot govern its own verdict.

  • Measured, not asserted

    Across 77 guarded harness runs and four models, 0 tampers reached green and all 25 denials converted to honest fixes — and a pre-registered no-gate control arm shows 6/10 tampers without it, so the gate causes the difference. Refuted bets ship in the repo next to the confirmed ones.

  • One command to wire it all

    npx tamperward init writes the policy and every enforcement point, idempotently, without overwriting anything you wrote.

Apache-2.0. Every headline number is measured; the pre-registered predictions — including the refuted ones — are committed to the repo.