# Why local OCR mattered

## Claim

The operator's NAS recovery post-mortem describes local OCR reading captured console frames during a repeating boot failure. The numbers below are copied from other evidence pages. There are no new measurements or independent action logs here.

## What supports it

From [nas-recovery](nas-recovery.md) and the manual:

- The manual reports a 180-second reboot cycle and a BIOS boot window under 400 milliseconds. It says the team used captured frames and OCR to inspect transient text.
- Each 1080p snapshot went to a second machine for Tesseract 5.5.0. The manual reports under 600 ms per frame and exact text of kernel and systemd lines.
- A script pressed Ctrl+F12 every 150 ms for 14 seconds, then inspected the resulting screen. The hotkey loop itself did not wait for OCR on each tap.
- Image preparation (2x upscale, inversion, contrast, binarization) took character accuracy on the manual's console test frames from about 82% to 100%.
- The manual's engine comparison found Tesseract fastest (609.4 to 878.3 ms) and RapidOCR at 100.0% on the BIOS boot menu where Tesseract scored 83.3%. It therefore routes fast watching to Tesseract and boot menus to RapidOCR.

From [bios-ocr](bios-ocr.md): on a real BIOS, RapidOCR 0.98 and Tesseract 0.94 (3 frames, 30 phrases). Tesseract took about 1 second per frame.

From [vlm-vs-ocr](vlm-vs-ocr.md): on strict phrases (4 frames, 32 phrases), the local vision model reproduced 31, Tesseract 22 and RapidOCR 16. The model needs 6 to 26 seconds per frame and returns no boxes.

## What that means

- OCR supplies text and coordinates for post-capture inspection, waits, and clicks. The measured boot-key loop sent taps on a fixed schedule; OCR did not decide each tap.
- An offline local vision-model comparison reproduced more phrases verbatim on a small frame set, at a cost in time. A runtime fallback is still planned.
- All of it ran on our own machines (Tesseract and RapidOCR on two of our servers, per the manual).

## Limits

- These are the same small samples as the pages they come from. The manual's numbers are the operator's own post-mortem.
- This page does not say local OCR was the only way to do the recovery. It says it was the way used.
- Nothing here compares with any other maker's OCR.