Discover
Scan the domain, signals and recurring friction. Look for consequential problems rather than attractive technologies.
Output: opportunity landscape and initial questionsEvidence-led engineering
Build evidence before building scale.
A practical method for moving from an ambiguous problem to an accountable decision. Architecture connects intent to implementation; evidence decides what happens next.
The position
An AI demo is not a product argument.
The familiar sequence - idea, LLM, demo, claim success - conceals the questions that matter. Is the problem worth solving? Is AI the right intervention? Can the architecture work inside its real environment? Do users understand the result? Are the controls proportionate to the consequence?
The method makes those questions explicit, produces evidence against them and treats stopping as a legitimate outcome.
Nine connected stages
From uncertainty to operation.
The stages are directional, not a rigid waterfall. Evidence can return the work to an earlier stage or end it altogether. Learning, revision, stopping and further experimentation are valid outcomes; not every inquiry needs to reach Ship.
Understand
Examine users, workflows, incentives, constraints, existing systems and failure costs. Separate the stated request from the problem underneath it.
Output: problem evidence, context and constraintsFrame
State the hypothesis, intended outcome, boundaries, assumptions and measures. Define what would falsify the idea before building momentum protects it.
Output: testable product and system frameArchitect
Define responsibilities, data flows, trust boundaries, human authority, governance, security and evolution paths. Use architecture to expose risk and enable change.
Output: architecture hypotheses and decision recordsPrototype
Build the smallest useful instrument that can answer a consequential question. A prototype exists to learn, not to impersonate production.
Output: observable behaviour and new evidenceProve
Challenge the riskiest claims: usefulness, intelligence quality, safety, performance, cost and comprehension. Seek disconfirming evidence, not only success cases.
Output: evaluation findings and confidence limitsEngineer
Turn what survived into durable software, data, interfaces and operational controls. Trace important requirements and decisions into implementation.
Output: production-oriented system and evidence trailEvaluate
Assess the complete socio-technical system against its requirements. Make residual uncertainty, limitations and operational responsibilities visible.
Output: system evaluation and release positionShip
Release deliberately with ownership, observability, support, safeguards and measures for the next learning loop. Shipping is a new evidence phase, not the end.
Output: controlled release and operating feedbackDecision gates
Progress is earned.
At major transitions, the current evidence is reviewed against the hypothesis, risk and value - not against enthusiasm or sunk cost.
Evidence supports the next investment.
The hypothesis remains credible, material risks are understood and the next stage can answer a worthwhile question.
The idea may be sound; the evidence is not.
Change the frame, architecture or experiment, then retest. Record what failed so the loop creates knowledge.
Stopping is a successful decision.
End work when the problem lacks value, the architecture cannot meet the need, evidence contradicts the hypothesis or adoption is implausible.
The questions at the gates
- Problem gate
- Is this problem real, consequential and worth solving?
- Architecture gate
- Can a feasible, governable architecture satisfy the need?
- Prototype gate
- Does observed behaviour support the core hypothesis?
- Evidence gate
- Do results justify further investment - or only expose a better question?
- Adoption gate
- Will somebody use it, trust it and sustain its operation?
- Release gate
- Is the complete system ready for its real consequence and context?