# 2025 Vumarii image — native-resolution transcription audit

## Source

- Account post: `1943302200622567872`, 10 July 2025, with the text
  `One thing @grok 4 cannot solve.`
- Native preserved media: `solution/tweets/media/Gvf_DZyXQAAcYYK.jpg`
- Dimensions: 853 x 892 pixels
- SHA-256: `e9ca65be6727620969e775c21ca0c0e46f6a8c498a85254d88e9f901515ec807`

The image is a 17-line Vumarii-family cursive text. It is not Page 23 and it
is not the 2025 factorisation square. The JPEG is a progressive X-hosted
re-encode; no higher-resolution original is presently preserved in this corpus.

## Predeclared transcription criterion

A complete **glyph-level** transcription would require, before attempting a
language decode:

1. every base glyph instance assigned a reproducible bounding box and visual
   family;
2. every detached dot, tick, comma, or slash attached to a base glyph or marked
   as independent under a fixed geometric rule; and
3. fewer than 1% unresolved or tied assignments after a second pass from the
same source bytes.

Word spaces are not part of the criterion: this later Vumarii register uses
*scriptio continua*, unlike the 2019 image's three-dot separators.

## Current result

The retained native-resolution working sequence
`solution/vumarii_long_mapped_s50.txt` records all 17 visual lines and 824
base-glyph tokens, mapped provisionally into the older 50-family catalogue.
It explicitly reports 166 low-confidence assignments (20.15%). Its `/` marks
are pipeline group boundaries, **not** observed word separators.

Therefore the image has a reproducible line-level working transcription, but
does **not** meet the predeclared standard for a complete glyph transcription.
The earlier claimed English alignment to a Scienceline article remains retracted
and is not used as a crib or validation target.

## Next valid operation

Build a fresh native-image segmentation table containing bounding boxes,
component masks, base-glyph family, attached-mark decision, and confidence for
every one of the 824 base tokens. Only then can repeated sequences and possible
language models be tested without retrofitting letter choices to desired prose.