Antfly Circus Maximus

Antfly parser evidence · reviewed page specimens

Parse
Atlas

Native PDF renders, extraction outcomes, and Reader evidence organized by retained run. The newest reviewed run opens first.

3 result sets pdf-hard · OHR-Bench · ParseBench

pdf-hard · 27 August 2026

pdf-hard · 308f221ea31d

Per-page differential evidence from pdf-hard: native Antfly rendering and Reader output beside the same call over a PDFium reference render.

Reviewed
Antfly
antfly dev (zig runtime) · PR #574
Commit
94604a468494
Binary
308f221ea31d1f31…
Run scope
12 / 749 pages
Atlas cards
12

pdf-hard was created around OfficeQA Pro Treasury Bulletin failures, which remain its primary defect case; OHR-Bench and other hard-page sources broaden the compatibility matrix.

Reference text is diagnostic. Pixel comparison and Reader behavior remain distinct evidence layers.

OHR-Bench pages are reproduced for research/non-commercial error analysis; public OfficeQA/FRASER pages retain their source terms.

Run triage

Render first. Reader second.

pdf-hard separates pixel divergence, extraction termination, and Reader limits so one failure class cannot hide another.

Blocking

11terminal or severely corrupt pagesThese pages cannot support a valid downstream OCR comparison.

Degraded

1retrievable but suspect pagesPixel or Reader evidence remains below the reviewed boundary.

Completed

1of 12 tested pages11 terminal/no-request outcomes

Render match

1perceptual PSNR ≥30 dB0 corrupt · 0 suspect

Selected pages by source group

academic11 blocking
administration65 blocking
manual22 blocking
news22 blocking
textbook11 blocking

OCR classes

The page class travels with the failure.

Counts can overlap because one page may be empty, numerically incomplete, and renderer-degraded.

4

Empty

Inventory class selected before native OCR execution.

1

Garbled

Inventory class selected before native OCR execution.

2

Low ground truth recall

Inventory class selected before native OCR execution.

6

Short

Inventory class selected before native OCR execution.

PDF feature families

Compatibility is a codec-and-font matrix.

Page-level filters describe the original PDF objects, not the exported atlas image.

Image/filter familySelected pages
/JBIG2Decode+/JPXDecode1
/JPXDecode11

Page evidence

12 selected pages

Static source images are metadata-free WebP renders; only the active tab and current card page load.

Loading page evidence…

Loading 12 page cards…

Repair queue

Fix the failure, not the counter.

Each recommendation ties an aggregate class to a product-level change.

Match the reference render

0 severe and 0 suspect native renders are retained in this run.

Add the retained page object and expected pixel behavior to the narrow native codec or renderer suite.

Separate render and Reader limits

1 pages lose at least 0.20 token recall through the native render.

Fix pixel divergence before changing OCR models or acceptance thresholds.