w = λ · κ

Status: a model with provisional parameters. A pre-registered test of the model has been designed and has not been run. The only data so far are one exploratory N-version study (NX, logbook tier C/D, inconclusive). The exposure λ is estimated from detector counts, normalized by code size; in NX it was the baseline defect rate of the generator. It depends on the model and the task, not only on the project.

Reads as

For one family of failures, w is the expected cost the project carries. λ is the exposure: how prone the project is to that failure. κ is the cost if the failure reaches production. Multiply them and you have the weight of that failure family for that project. λ is a density, so w is expected cost per unit of code and per period, not the project’s total.

What it is for

It ranks failure families by what they are worth to a project, so that you spend checking effort where the expected cost is highest. It is the first step of the revival model, which asks which long-standing engineering practices become worth reviving when an AI executor makes them cheap.

What to keep in mind

  • λ is conditioned on the generator. In the revival work, the exposure to a failure depended on the model and on the difficulty of the task, not just on the project. The model is an argument of λ.
  • Not measured means not counted. If λ or κ cannot be determined, the model records it as not measured and lists it as unknown. It does not guess a penalty.
  • The test has not been run. The revival program is pre-registered: the design was frozen before any generation, so a reader can check a future result against it. The grid itself has not run. See the pre-registered design.

Where to read more