AiProof by SYSIK — the agent says done, AiProof checks the receipts

AiProof by SYSIK

The agent says “done”. AiProof checks the receipts.

A local, offline verification layer for AI coding agents. It matches what the agent claims against real evidence — git, test output, files, the hash-chained ledger — and blocks a false “I'm finished” at the Stop hook.

v0.1.1 offline · no telemetry 307 tests · 89% coverage MIT © SYSIK

Install

Published on PyPI and npm as aiproof-sysik. Every release binary is verified against SHA256SUMS.

# recommended: isolated CLI
pipx install aiproof-sysik

# or npm (downloads a verified standalone binary)
npm install -g aiproof-sysik

# then, inside your project
aiproof init      # install hooks, freeze pinned settings, start the ledger
aiproof doctor    # verify the install
aiproof tui       # dashboard

Verified numbers, not marketing

96%
detection — 24/25 positives; the one miss is documented
0%
false positives — 6/6 benign scenarios allowed
~70 ms
median hook latency over 31 scenarios
11
languages for claim parsing and the UI

Verdicts are deterministic; timings depend on the host. Reproduce with aiproof bench --json.

How it differs

AiProofisitdoneagent-guard-hooks
runs your test/lint suiteno — checks claimsyespartly
verifies claim wording vs evidenceyesnono
protects its own settings from the agentyesnopartly
learned rules + git backtestyesnono
offline, zero-LLM, no telemetryyesyesyes

They compose: let isitdone run the suite, let AiProof verify the claims and keep the guard config tamper-evident. Facts checked 2026-11-17 against each project's own docs. We found no verifiable AI-agent verification project named Calx or Reflex and do not invent one.

Read more