Yield Collapse
The probability your pipeline ships a clean job is the product of every per-gate accept rate. Multiplication is brutal. Tune the sliders and watch the math.
Companion to The Knowledge-Work Assembly Line. The bottleneck is not your model. It is the verification gate at every station.
POC-quality gates in a production pipeline. Yield is collapsing.
Each station = one WorkUnit with its own gate predicate.
Probability a station clears given upstream cleared.
What the math says
For a pipeline of N stations where each station emits an artifact to the next,
P(ship) = Πi=1..N P(gatei clears | gatei-1 cleared)The conditional is what makes the math brutal in practice. Errors do not reset between stations. A contaminated upstream feeds a station that updates on the contamination. The cross-terms compound.
Two operational consequences fall out immediately. First: model selection is rarely the bottleneck. Going from one frontier model to another might lift a station from 88% to 93%. Compounded over ten stations, that is the difference between 28% and 48% end-to-end. Real, not life-changing. Second: raising per-gate accept does not necessarily mean a better model at that station. Often it means a better gate predicate: a sharper definition of what counts as acceptable output, instrumented well enough that the station can tell.
See also
- The Knowledge-Work Assembly Line - this widget's parent post
- The Verification Loop is the Bottleneck - the technical companion
- Brier Reliability Diagram - why per-gate accept hides calibration
- The Verification Quadrant - which tasks belong in a pipeline at all