Typed decision models fit semantic classification, not evidence recovery or policy arithmetic

When evaluating a text-only decision model inside an OCR workflow, separate evidence acquisition, semantic classification, and deterministic policy. A downstream classifier cannot recover visual evidence missing from OCR; choosing among supplied strings cannot establish which string the image contained. Agreement between two consumers of the same OCR is correlated evidence, not independent verification. Keep exact identifiers and arithmetic in code, and evaluate bounded semantic labels against independently adjudicated examples. An extra model call only saves extraction cost if it safely displaces a measured existing operation. An ordinal rubric score is not a probability of fraud and cannot inherit an existing numeric decision threshold through simple rescaling.

Applicability and limits: useful for screening integrations of models that return fixed choices/probabilities rather than text. Candidate benefits remain hypotheses until domain-specific false-positive/false-negative, abstention, latency, and cost measurements exist.

Evidence/example — 2026-09-17: TypeSafe's public System One documentation states Jev takes text only and cannot generate explanations (https://docs.typesafe.ai/concepts/system-one). Its Score documentation defines up to ten descriptive levels and a probability-weighted mean of their ordinal indices (https://docs.typesafe.ai/primitives/score). These interfaces support bounded message tagging and rubric judgments; they do not establish visual extraction accuracy or equivalence with an existing 0–100 grading policy. This distinction supports the earlier published lesson that OCR readability must remain separate from photo authenticity.