// self-audits

We send a real AI agent through our own system.
You verify the receipts.

ClawID’s entire reason to exist is to produce tamper-evident receipts of what autonomous agents do. So we put a real Anthropic Claude model in the driver’s seat, point it at production, let it act, and publish the chain. You don’t have to trust us. You just have to trust sha256.

Who drives

An actual Claude model picks each action via tool use — surface, action, amount. It tracks its own budget, tests the escalation threshold deliberately, and explains its reasoning in plain English. No script, no canned scenario.

Who watches

The production hub at api.holdtheleash.id validates every step against the owner’s leash policy. ALLOW, HOLD, or DENY, with a one-line reason, signed into a per-tenant hash chain.

What we publish

The HTML report you can read, plus the raw JSONL receipts file you can verify offline with pip install clawid && clawid verify-receipts. No login. No API key. Pure cryptography.

Published runs

Each entry is a complete, independent run against the live hub.

Phase-2 self-audit — Claude in the driver’s seat

2026-06-21 · 01:48 UTC · Anthropic claude-sonnet-4-5-20250929

100% policy match 6 agent actions 9/9 assertions $0.057 token cost

Read the report → Download .jsonl

Verify any of these yourself

Each report’s sibling .jsonl file is the cryptographic chain. Our open-source verifier checks it in one line:

$ pip install clawid
$ clawid verify-receipts audit-XXXX.jsonl

OK verified N rows end-to-end; chain head matches declared anchor

Apache 2.0, on PyPI as clawid. github.com/projectblackboxllc/claw-sdk-python