Detect when the plant is outside the training distribution and defer before acting.
The AI-Native S.M.A.R.T. Generator Master Blueprint — eight layers (L0→L7), one control stack, wired to both machines. Telemetry rises in microseconds; control descends the same path.
Category: A · methodology · Plugs into: L3 · Horizon: FOAK · Status: on the roadmap — not yet built
What it is
A dedicated out-of-distribution detector flags plant states unlike anything seen in training, and separates epistemic uncertainty (the model does not know) from aleatoric uncertainty (irreducible noise).
The method
Density models, deep ensembles, or Mahalanobis distance in feature space; high epistemic uncertainty routes the decision to rules or to a human instead of to an actuator.
Why it matters
The most dangerous AI failure is confident wrongness off-distribution; this makes the loop degrade toward caution exactly when it should. Plugs into L3, feeding the L4 gate.