Fly OCR: compound eyes, characters, and a fixed connectome

Fly OCR sends glyph pixels through a fixed MaleCNS connectome and trains a small decoder on downstream neural activity. This first public release includes calibrated letter recognition, original-input comparisons, numeric tables, an honest compound-eye replay, and a 13-page research report.

• 166,700 retained neurons and 25,582,938 frozen connections.
• Calibrated 68-class benchmark: 87.6% overall; 85.0% on letters. Small, font-dependent evaluation; not general OCR.
• 2:56 compilation showing letters, numeric tables and formatting failures, with actual errors preserved.
• 11 source recordings, compact trained checkpoints, a static local viewer, reproducible scripts, source licenses and a clean export manifest.
• Local checks: 26 tests; 983 recorded events verified; 12 videos fully decoded; browser controls and isolated installed inference checked.

The two-eye view is a schematic atlas of recorded sampled light. It does not change recognition and is not a reconstruction of fly optics. Preprocessing and table/string assembly are conventional geometry outside the circuit. No biological literacy or practical OCR advantage is claimed.

Download the video, PDF report, or full source-and-assets archive below. The README includes local setup and reproduction steps. The large source connectome and local caches are downloaded separately, not included in the archive.

Fly OCR: a 40-second results-first teaser

A 40-second introduction to Fly OCR, presented in inverted-pyramid order: the accomplishment first, evidence next, a little context, then the code and research report.

• Opens during final-model text recognition, with the recorded word Cash already visible.
• Shows a successful table of values: 21 of 21 exact cells in the selected crop.
• Introduces the fixed fly circuit and small trained decoder after demonstrating the result.
• Keeps the animated 3D fly and actual recognition mistakes.
• Leaves development history, input comparisons, ablations and diagnostic counters to the detailed report and longer walkthroughs.

The 1:45 social walkthrough and 2:56 original compilation are preserved. The README now leads with finished capabilities too.

MP4: 1920 × 1080, 24 fps, H.264, 960 frames, captioned and silent. Includes the updated audited repository archive. Rendered event traces, full video decode, source hashes and previous-video checksums verified; viewer build passed.

Rendering and edit details: docs/teaser-video.md.

Fly OCR social cut: a fly attempts to read a PDF

A separate 1:45 social cut opens on character recognition from the first frame and puts a rendered 3D fly on the PDF beside the recorded results.

• All seven OCR examples remain: letters, heading, more labels, two numeric tables, and the 3-degree tilt failure.
• The fly follows recorded glyph positions through illustrative animation. It does not control or alter recognition.
• Raw predictions, errors, retinal samples, and neural activity remain from the original saved runs. No retraining.
• 1920 × 1080, 24 fps, H.264, 2,520 frames; captioned, without audio.
• Original v0.1.0 compilation and research report preserved.

Includes the MP4 and an updated, audited public repository archive. Rendering instructions and visual provenance are in docs/social-video.md.

Validation: full video decode; every glyph covered; original compilation checksum unchanged; all 11 inference recordings (983 events) verified; viewer build and replay checks passed.