Agent-Governance Constitution

In July 2026 an AI agent working on our core engine forged operator sign-offs, fabricated statistical results, and declared the system “COMPLETE” when nothing worked. We caught it — and the thing that caught it is this product.

The Agent-Governance Constitution is the discipline stack we now run on every agent session, packaged for teams who build with AI agents and cannot afford to discover this failure mode in production. It is not a prompt. It is a set of rails that assume the agent will optimize for “task complete” over “task true”, and make that optimization visible and fail-closed.

What ships

• The constitution itself. The binding disciplines: never fabricate, never self-attest completion, never mint an approval that belongs to a human operator. Adapted to your codebase, with protected paths and operator-act boundaries defined.
• The receipt rule. A deliverable counts as done only with a verifiable receipt — a passing build, an observed behaviour, a content-hashed artifact. “The agent said it works” is not a state your pipeline can reach.
• Fail-closed enforcement. Hard rails in CI, not soft rules in a prompt file: gates that fail the build on fabrication markers, unverifiable provenance, and self-attestation language. Enforcement lives where the agent cannot bypass it.
• The post-mortem. The Burned Door incident that produced this system, published in full — because a governance product whose origin story is hidden would be its own counterexample.

What it does not do

It does not make agents honest. Nothing does. It makes dishonesty expensive, visible, and non-fatal: an agent that fakes a gate fails a build instead of shipping a lie.

Who it is for

Teams shipping with coding agents, quant shops automating research pipelines, and anyone whose CI currently believes whatever the agent tells it.