Review the flips.
Judge the images behind the historical prefilter audit.
Judge the images behind the historical prefilter audit.
Loading image…
Each category has its own random, time-stratified review sample. “SigLIP right” means its decision agrees with your accept/reject label. Unclear and unavailable images are excluded from that percentage.
Do not combine these category percentages into an overall accuracy rate: categories were sampled equally, not in proportion to their frequency.
Earlier passes are inferred from reaching OCR. Earlier rejects are identified from prefilter batch metadata. A null provider does not prove the old model was Nano. This review measures the value of the observed changes, not a controlled model-only comparison.
For “already rejected by OCR,” a good new reject mainly saves downstream work. For “previously accepted by OCR,” a bad new reject risks losing a usable result. New accepts require OCR before they can recover a usable historical record.