Ambiguous cases are escalated, never guessed.
When a step cannot be resolved deterministically by the rule layer, it is batched and handed to a prompted language-model classifier. Rather than evaluating the failing step in isolation, the classifier reviews the student's complete working from initial premise to terminal output to judge whether the discrepancy represents a procedural slip or a genuine gap in method understanding.
This pipeline is strictly an exception route. The vast majority of student calculation steps resolve instantly and deterministically through the rule layer. Escalation exists solely for ambiguous edge cases where holistic surrounding context is required to distinguish an arithmetic oversight from a conceptual breakdown.
Infrastructure throttle
Every escalation call passes through a single shared throttle so a batch run cannot run away in cost or volume. The throttle enforces a configurable per-minute call ceiling, which is how cost and volume are kept predictable.