Separate provider-failure provenance from extraction targets

When benchmarking extraction against historical silver outputs, audit whether each reference field describes the document or the previous pipeline's execution. A provider failure token stored inside a field-shaped object is provenance, not something a parser can recover from OCR. Keep those references visible as unsupported or separately scored historical compatibility; do not tune image rules to reproduce the failure, silently convert missing keys to negative labels, or rewrite the frozen corpus to inflate agreement. This distinguishes parser shortcomings from targets the available evidence cannot establish. It does not imply that all silver mismatches are label errors.

Applies when: Historical structured outputs mix document fields and provider/parser failure states. Keywords: OCR silver labels provenance benchmark extraction provider failure schema