Freeze cross-workflow confidence gates before OCR crop benchmarks

Observed 2026-09-08 while designing Shield COR-4083 from the preserved 2026-09-04 Gemini/RF-DETR benchmark contracts: when model outputs expose confidence but only some workflows have a numeric production floor, predeclare any added threshold as experiment-only before opening the cohort and keep it distinct from production policy. Also report same-model original-versus-crop equality as agreement, not accuracy; an independent source review is required before naming accuracy. This prevents cohort tuning and jointly-wrong outputs from inflating results. Provenance: repository OCR contracts plus preserved benchmark runners; no production records or private values included.