Full confirmation is a defensible scaffold and a poor destination. Asking a learner to review every word preserves control but makes access feel like extra work.

Academic risk is more concentrated. Numbers, operations, negation, quantities, entities, direction, conditions, and technical fingerspelling can change the answer even when the rest of the sentence is acceptable.

The current selective-intervention layer recovered all 34 predeclared critical checks across 24 sealed pairs and produced zero unsafe ready states.

The product implication is precise: surface the smallest decision that protects the lesson, keep the route closed until the signer resolves it, and automate only when independent calibration earns that privilege.