Why I built the governance layer

2026-09-11. The operating thesis behind this product.

Every support operation I have run hit the same wall. Governance was side work. Nobody owned quality, access, or cleanup, so quality rotted, access sprawled, and cleanup never shipped. Then something broke at 2 a.m. and the runbook turned out to live in one person’s head. I decided to build the layer I kept asking vendors for and never got.

The requirements were non-negotiable. Every write screened before it lands. Every definition owned, one meaning per thing. Every recall carrying its lineage, so a wrong answer traces to the exact row it came from. Access scoped per integration, so each tool sees what it needs and nothing else. Deletion that produces a certificate instead of a promise. An audit chain that answers who decided something was true without scheduling a meeting.

This is warehouse discipline applied to agent memory. Screened writes. Owned definitions. Checked lineage. Scoped access. Proof. The underlying systems differ, but the governance pattern is identical, and it is the pattern most AI deployments are missing. A definition nobody owns drifts. A write nobody checks poisons everything downstream. A deletion nobody can prove is a liability with a date on it.

My bench is Zed plus OpenCode, with Claude Code for heavy lifts and OpenClaw running production automation. The OpenClaw integration is deliberate, not incidental: per-turn recall inside an untrusted fence, writes gated as proposals, merge-seam forgery stripping, all covered in the stateful prompt injection piece. That loop builds monitors and triage tooling daily. The rules live in code, where they cannot be forgotten, skipped under pressure, or rubber-stamped at end of day. When a write is refused, the rule is written down and the tooling makes approval easy. Make the right action the cheap action and a small team covers ground that used to need a large one.

This repository is the evidence. Every control named above runs here, pinned by tests, with residual limits stated where a buyer can price them. If you are evaluating this for a regulated operation, start with the guarantees in the README, then ask what broke to earn each one. There is a drill record behind every answer.