Harness engineering equips a run; loop engineering drives runs; governance engineering is what keeps the loops honest: state authority, say/do reconciliation, and rule lifecycle.
This repository is a disclosure package for governance engineering as a built system — an execution-governance layer for autonomous LLM agents, built and used in private between May and July 2026. It documents the design thinking (essays/), the mechanism specifications (specs/), and a verifiable development chronology (CHRONOLOGY.md). Implementation code is deliberately withheld (DISCLOSURE-MAP.md explains the tiering).
Prompt engineering shaped what a model is asked. Context engineering shaped what it knows while answering. Harness engineering shaped the machinery an agent acts through. Loop engineering shaped how long-running work continues across turns. Above all of these sits an unshaped layer: what makes an autonomous run trustworthy — not whether the agent CAN do the work, but whether what it claims about the work is true, whether its plan survived contact with reality, and whether the rules governing it are themselves governed.
This package documents one worked answer, built as a running system rather than a proposal:
- Two blocking audit gates — a plan is audited by an isolated-context auditor before execution; execution is audited before "done" is allowed (specs/gates.md). Enforced at runtime by hooks, not by asking nicely.
- Say-do reconciliation — six machine checks that diff what the run's state CLAIMS against what its artifacts SHOW (specs/say-do-checks.md).
- A four-cell deviation taxonomy — every checkpoint can classify what the plan met: on-plan, detour, grind, or escalate — classified by the response the deviation licenses (specs/four-cell-taxonomy.md).
- Evidence-graded state — claims carry their anchor quality (weak / medium / strong); completion requires medium or better (specs/run-state.md).
- A self-taxing rule lifecycle — every rule and mechanism must pay rent; promotion needs an owner's tick, doubtful machinery goes on probation, dead rules get retired (specs/rule-lifecycle.md).
The package's value rests entirely on claim credibility, so CHRONOLOGY.md holds itself to an uncomfortable standard: every date is "existed no later than", every claim carries a graded citation, each anchor's grade is stated where it is used (four claims carry externally held S-grade anchors; everything else is local evidence, and says so), and claims that only the owner can attest are quarantined in their own section. NEEDS-OWNER.md lists the evidence that would upgrade the rest.
Prose (essays, chronology, specs, this README): CC BY-SA 4.0 — see docs/LICENSE-DOCS.md. Code and skill files (none staged at initial publication; implementation is withheld per DISCLOSURE-MAP.md): AGPL-3.0-or-later — see LICENSE.
The name (governance engineering) and licenses are settled owner rulings; any release of withheld implementation remains a per-item owner decision. Contributions are not being accepted until after first publication.