AI is the
capability —
and part of
the team
delivering it.
Two dissimilar intelligences, each positioned to catch the other's failure modes. This is the engineering discipline for pairing them — and it is free to read, question, and use.
Agents and humans bring different, changing strengths and failure modes that must be measured locally rather than assumed forever. Under the present boundary, humans retain recognized accountability; capability, ethical performance, moral status, and authority remain separate questions. The framework is the discipline of pairing them—and bridging novelty when neither begins with enough evidence.
Fourteen principles that survive all tailoring
Each states a belief the framework will not bend. Read them in descent, or jump to any one from the index. Watch each demonstration — every motion here is the principle, not decoration.
The principles are the surface. This is the depth.
Seven parts, forty-two chapters, organized so you can feel the architecture without reading it linearly. A visitor who read only the fourteen movements already has the whole argument; everything here is for when you want to operate it.
Enter through the summary. Open only the depth you need.
Choose a part and chapter. Highlights remain visible; governing boundaries, evidence, collaboration guidance, failure modes, and inspection prompts open progressively.
Embedded, non-rendered: the complete authoritative agent corpus for machine consumption.
The lifecycle & its gates
Eight stages, each ending in a decision gate that releases funding on evidence — not on the calendar. Select a stage; see what the gate demands and what it commits.
The teaming patterns
The framework's most portable contribution: a pattern language for pairing, each pattern named by the failure mode it covers. The bracket ⌐¬ means "each covers the other" — never a merge.
The automation spectrum
Work is decomposed to the decision, not the job. Every decision gets an allocation — including the prohibition class: work that should not be automated, with the reason recorded.
Consequence & transition tempo
Two coupled dials. Consequence sets how much framework you need; tempo sets how fast you move. Rapid is forbidden above a consequence ceiling — and the reversibility ratchet can force you slower still.
The survives-the-burst test
A counter-cyclical filter run at the funding gates. It starves hype-driven bets before a crash and keeps funding evidence-backed winners through the trough. All three predicates must hold. Toggle them.
Never let capital run ahead of evidence.
Bubble exposure is a managed risk, not a forecast to get right. The test is the machine-checkable half of that discipline.
The eleven properties
A conjunctive mandate — all eleven must hold together. Not a score, not a ranking, not a progress meter. Hover any one: the rest dim, because the failure of any single property can invalidate the capability.
The anti-patterns catalog
Twenty-two ways delivery fails. The catalog doubles as the Sentinel's monitoring specification: where a failure is machine-detectable, it is tagged sentinel. Select any to open it.
If you read nothing else
Deliver AI capability the way you would deliver anything whose value is uncertain, whose cost outlives its launch, and whose behavior drifts — but do it with AI on the team, not just under the microscope. Move money when evidence arrives. Scale rigor to consequence. Design the boundary between human and machine so it is observable and reversible. Point oversight both ways. Keep accountability human. Name the workforce cost out loud. And never let capital run ahead of evidence.
Principles ten through fourteen are the framework's distinctive claims — the arrival, not the coda.