The handoff between Classifier and Recognizer
The two modules do not call each other. They are coupled via artifacts: the Classifier leaves behind path and metadata, the Recognizer reads them and infers its model from them. This loose coupling is why partial runs, repetitions, and corrections can all resume at the same point.
Classifier
via artifacts
module call
Recognizer
Model covers the category
The standard case: analysis starts immediately with the assigned model and its engine. No intervention, no waiting state.
Category trained, model missing
The automation creates a recognition model itself: field schema suggested from the triggering document, analyzer published, the document attached as a permanent example. The target system assignment remains intentionally open — unverified data never reaches the target system. At most once per category.
With reason, not silently
If no model remains, the case is written to the journal with a plain-language reason — “no recognition model for classifier X” or “category Y not configured in model Z”. A document never disappears.
Each recognition model carries its engine: classic trained field extraction or an authorized field schema. The Recognizer branches per model — everything after that is engine-neutral because both paths produce the same result format. The Classifier remains completely untouched by this.