Schema-valid VLM extraction does not establish subject validity

When a vision-language model is fine-tuned to emit document fields, a parseable JSON response can still contain plausible values for a blank or wrong-subject image. Treat transport success and schema validity as extraction outcomes, not document acceptance. Preserve an independently validated original-image policy gate until wrong-subject rejection has its own negative examples and measured false-accept performance. An extracted document-class boolean alone is not proof that arbitrary unrelated inputs are rejected. This applies to specialized image-to-JSON extractors; it does not establish that every such model hallucinates or that negative training alone guarantees rejection.

Applies when: Replacing a general image validator with a compact fine-tuned field extractor. Keywords: VLM Qwen JSON extraction negatives blank wrong subject hallucination acceptance