Run Date: 2026-09-09 22:10:34.945927 Copyright (c) Citizens' Oversight
| setting | value | note |
|---|---|---|
| vendor | 'ES&S' | |
| ml_min_detection_probability_pct | 96 | positions scoring below this are never detected at all |
| grid_rows_to_drop | 2 | last N y-grid rows excluded from candidate positions entirely -- no in-code default, vendor-specific unless a job overrides it |
| ml_target_exclude_regions | none | |
| ml_target_eval_w x ml_target_eval_h | 72 x 32 | classifier's own evaluation-clip size |
| target_offset_x/y (page 1) | 0, 0 | |
| target_offset_x_p1/y_p1 (page 2+) | 0, 0 | |
| mapping_ocr_detect_only | False | |
| mapping_ocr_match_only | True | map_option_names_ocr()'s own two-flag split -- affects whether map/ml_detected_targets.csv reflects a real detection pass or a reused one, not this report's own generation |
1 of 1 styles have detected oval count exactly matching the EIF-expected count.
No mismatches.
Full snapgrid (every 'top'/'lft' line, edge to edge) in thin red, and the oval classifier's own evaluation-clip boundary at EVERY candidate grid position (not just accepted detections, keepout regions already excluded) in thin green -- lets you check grid registration against the page's real timing marks and spot an eval window overlapping neighboring ballot content directly. On top of that: for every style with detections in map/ml_detected_targets.csv, a bold red box + 'x,y,confidence%' around every detected oval, a bold blue box around the region that will be sent to the AI/OCR matching stage, and a bold green box previewing the eventual crop boundary. 2 images across 1 styles.
01-mldetect-redlined-1.png
01-mldetect-redlined-2.png
Run Date: 2026-09-09 22:10:34.946039 Copyright (c) Citizens' Oversight