09 METHOD
Two mechanisms do the work. Four gates decide what gets built, with the pass bar written before any test runs. An agent studio does the building, under one rule that never bends: agents draft, a human ships.
09.1 THE FOUR GATES
Every idea, including the founder's favourites, walks the same bench. It can be sent back or parked at any gate, and parked work stays on the public record.
Seven questions in parallel: is the problem real, who has it, has someone already solved it, can it be built and run lawfully, will anyone pay, what's the return for the effort. Then one reviewer whose only job is to kill it.
The riskiest assumption gets one cheap, honest, real-world test: a fake door, a hand-made concierge report, an ad probe. Budget capped. Metric named. Pass bar locked before launch and never moved after.
The agent studio builds to a written spec: product, design, engineering, QA, security review, release. Work-in-progress limits keep it to one lead product at a time. Ships when it's true, not when it's loud.
A dated ledger of what real people did. Retention over signups. Watch-metrics published up front so anyone can hold the workshop to them. Weekly numbers feed the next decision: scale, iterate, park, sell.
09.2 THE STUDIO
The studio is a set of coordinated AI teams, each with a lead the founder talks to, specialists that do the work, and a critic whose job is to say "not good enough". Every team ends at the same place: a human decision.
129 AGENTS ACROSS THE TEAMS AS AT SEPTEMBER 2026 · REVIEW BUDGET THROTTLED SO DRAFTS NEVER OUTRUN THE HUMAN
09.3 THE RULES
A test's pass mark is written down before it runs. If the number comes back short, it failed. There is no "but the signal looks promising".
Every team carries a critic. In validation it is a red team whose only success is killing the idea. The best blueprint the process has seen was parked anyway.
A real practice uses every product before an outsider does. Feedback arrives on Tuesday, not next quarter. Nothing is sold that hasn't been lived with.
Clinical data lives in clinical systems. Products hold the minimum, and where a tool can run with no data at all, it does.
Work-in-progress limits stop four half-builds. The portfolio names one horse; the rest are sequenced or parked, in writing.
Parked and killed work stays published with its score and its reason. The ledger is what makes the wins believable.