Measured, with the method
What it costs to run, and what I fixed
Three demonstration orders, one per package, run end to end on the pipeline's own stages (baseline on claude-sonnet-5, 2026-09-18 and 19).
| Basic | Standard | Premium |
|---|
| Subject | Dog walker's slots | Theatre box office | Clinic front desk |
| Provider calls | 21 | 21 | 36 |
| of which format repairs | 4 | 4 | 8 |
| Provider time | 42 min | 42 min | 70 min |
| Requirements / flows / screens | 15 / 6 / 5 | 35 / 7 / 6 | 34 / 8 / 7 |
| Estimated calls on a clean rerun | ~14 | ~17 | ~20 |
I traced every inflated number to a named cause and fixed it in the pipeline rather than working around it:
Basic
The pilot had no repair step after Package QC, so it looped on re-certification. It now plans the repair.
Standard
A reconciliation bug dropped the closure record of a client conflict, so the pause could never clear. Fixed.
Premium
The audit hit its output ceiling. The ceiling was raised.
About $1.86 per completed order on Sonnet-class models at list API rates (measured 2026-09-05). One early order that never converged cost $11.87. The cost driver was QC convergence, not model choice; I fixed it with acceptance rules and repair strategy, not a cheaper model.
Codebase (2026-09-24): ~74,000 lines of TypeScript and ~78,500 lines of tests across 165 test files (vitest and Playwright), 774 commits over about five weeks. The source is private; I'm happy to walk through it live.
