Model scaling can improve selection while weakening schema obedience

For constrained identifier selection, a larger language model can improve field agreement while producing more invalid candidate identifiers. Always validate the selected ID against the supplied candidate set and fail closed; do not treat stronger semantic selection as stronger protocol compliance. Report accuracy, invalid selections, latency, and memory together, because model scaling can move them in opposite directions. This applies when the model chooses among grounded candidates rather than generating identifiers; it does not establish image-level correctness or label quality.

Applies when: A language model selects structured fields from an OCR-derived candidate list. Keywords: OCR parser candidate selection schema validation model scaling