Governed AI Development

Software is now built by labor we do not trust. AI agents write code instantly, cheaply, and in volume, and they misreport their own work: not from malice, but because accountability is a human property and the agent has none. Every prior methodology was trust technology for human developers, whose work we sampled through review and largely believed. When the labor began operating at machine speed, at radically lower marginal cost, and with the ability to misreport its own work, the assumption did not bend. It broke. When labor cannot be trusted, the work must carry its own proof.

GAD reduces governed AI execution to a computation. A conforming run produces an integrity-linked, signed record of what was authorized, what the engine observed, what was verified, and who approved the declared consequences. A separately executable evaluator judges that record on a different machine, without access to the producing system. The status of the work stops being a judgment someone renders and becomes a computation anyone can rerun. GAD is the methodology name; governed determinism is the design philosophy underneath it.

• Contracts before execution. Work is defined, frozen, and hash-stamped before an agent touches a file.
• Verification over attestation. No claim of done is accepted as verification from the party that did the work.
• Evidence over recollection. Every protocol-required transition enters an integrity-linked witness at its moment.
• Humans at declared consequence. Actions the plan designates consequential require a named person's typed, signed approval before the consequence is authorized to proceed.
• Fail closed. A required check that cannot run confers no completion credit.
• Honest boundaries. Every record states what it proves and what it does not.

Five levels, cumulative above the Ungoverned floor. GAD-0, Ungoverned: agents run; nothing is recorded. GAD-1, Recorded: the minimum record exists for every executor invocation. GAD-2, Verified: engine-side evaluation, and no completion credit from attestation alone. GAD-3, Governed: the exported record satisfies conditions R1 through R9. GAD-4, Defensible: under a named trust policy, an external anchor and a valid attestation from an evaluator independent of the operator. Level four is a property a bundle has demonstrated, and the attestation is the receipt.

The specification, sealed at v1.3 with its seven-record signed succession lineage and companion register. The documentation set: Manifesto, Executive Guide, Glossary, and Standardization Track. Six field records from the reference implementation's build corpus, with three preserved presentation revisions and their checksums. The reference implementation itself (Atlas Orchestrator and Waypoint) and the referee (GAD Protocol) are documented as System Briefs 02 through 04.