Practical consequences
- Takt is set explicitly and low — on the order of 3 to 5 certified units per week at current staffing. Every WIP limit derives from it via Little’s Law.
- The only way to raise plant capacity is verification leverage, not more agents: reference-free judges that turn human review from “check everything” into “audit the auditor.” Every point of judge-alignment gained is capacity gained. This is Thesis 1 seen from the demand side.
- Overproduction is redefined. An agent built faster than it can be verified is not early; it is waste rusting in the yard. The classic first waste — building agents nobody asked for — gets a sharper sibling: building agents faster than anyone can trust them.
The design of the two touches
The stations exist to protect the attention budget. Everything mechanizable is mechanized; what reaches the operator is judgment that only the operator can supply. At Station 0 that judgment is intent — what should exist and why. At Station 4 it is acceptance — whether the claim and its evidence deserve a signature. What the operator reviews at certification is the claim and its evidence, not the work — that is what the verification shop bought. Automating everything except the decisions that matter is not a compromise. It is the point of the whole system.Self-audit
| Claim | Source |
|---|---|
| The customer is the operator; takt is his sustainable absorption rate for verification and intent | docs/factory-design.md:71 |
| One-human plant; attention twice per unit (intake, certification); trust-bandwidth is the constraint | docs/factory-design.md:71 |
| 3–5 certified units/week; WIP limits via Little’s Law | docs/factory-design.md:75 |
| Verification leverage is the only capacity lever; audit the auditor | docs/factory-design.md:76 |
| Overproduction redefined: faster than anyone can trust them | docs/factory-design.md:77 |
| Human reviews the claim and its evidence, not the work | docs/factory-design.md:146 |