Poka-yoke.
Japanese manufacturing term (ポカヨケ) for a mechanism that makes it physically impossible to produce or ship a defective part.
How it’s computed
Not a metric. A principle. The original examples were physical: a fixture that only accepts a part in the right orientation; a light that illuminates when a required step is missed; a checklist that cannot be signed off until every item is ticked.
The common thread: the system prevents the error, rather than relying on the operator’s attention.
Worked example
The classic poka-yoke example is a USB-C connector — it is reversible, so you cannot plug it in the “wrong” way. An older one: a gas pump nozzle that physically does not fit into a diesel tank. The error is made impossible rather than warned against.
How Yoke Agent uses it
This is the product’s name and philosophy. Yoke Agent is a jig for AI engineering — it makes it impossible to ship a configuration worse than the one you measured, because every claim about a configuration is backed by a reproducible grid-search and a numeric score.
The human-in-the-loop approval step at every phase (dataset, grid, report) is another poka-yoke: you cannot accidentally skip review.
Frequently asked
Who coined the term?
Shigeo Shingo, a Japanese industrial engineer, at Toyota in the 1960s as part of the Toyota Production System. The original term was baka-yoke (“fool-proofing”) but was softened to poka-yoke (“mistake-proofing”) because workers objected to being called fools.
How does poka-yoke translate to AI?
Evaluation is the poka-yoke for AI systems. A measured score, backed by a reproducible sweep, makes it impossible to accidentally ship a worse configuration than the one you decided on. The jig is the grid-search plus the approval workflow.
Other examples of poka-yoke in software?
Type systems that prevent mismatched function calls at compile time. Database migrations that refuse to run if the schema is out of sync. Git hooks that block commits with failing tests. All are poka-yoke by a different name.