Skip to main content
A production line is not a straight path; it is a bundle of loops running at different frequencies. The stations move units forward; the loops keep the whole plant true. Each loop declares the artifact that proves it ran — because per Thesis 3, a loop that stops looping is itself a defect, and the watchdog treats a missing loop-artifact exactly like a missing report.

​
The loop catalogue

​
Per-run learning and graduation

Row 12 of the compilation table lands here, at Station 6: the learning-file convention plus a weekly graduation job that proposes germline PRs. A lesson that appears across multiple run files gets baked into the standard work under full germline gates, and the learning files that taught it are retired. Insight flows upward on a schedule, not by accident.

​
The obeya — one screen, four families of number

Row 15: the metrics board plus the decision log, fed by certificates, queue, and telemetry:
  • Flow: WIP by station, queue age, lead time from work order to certificate, takt adherence
  • Quality: first-pass yield, escaped defects (baseline zero — keep it), judge/human alignment rate — the trust metric that buys capacity
  • Corrosion: stale-certificate count, oldest unmet re-certification, silence alarms fired and resolved
  • Economics: cost per certified unit against target, reuse rate, kaizen-budget consumption — 10% of compute, actually spent
  • Banned: lines of code, issue count, commit count, hours — vanity, by name

​
WIP limits and takt

Row 16: board configuration plus an alert when WIP exceeds the Little’s-Law bound derived from takt. Takt is set explicitly and low — 3 to 5 certified units per week — because the constraint is the operator’s attention. The loop here is the pacing itself: a full cycle at takt with zero silent stalls is the plant’s definition of running at rate — every gap in flow produces an alarm, and every alarm produces a response.

​
Self-audit