# Brief given to each classifying agent (verbatim, with the batch file path filled in)

You are classifying image crops from scanned printed pages. Each crop file shows the same region
twice, stacked: on top, the plain scan; below, the same scan where ink has been coloured. Blue-grey
ink is ink that a machine text layer has a word box over. RED ink is ink that no word box of the text
layer covers. A vertical strip of the crop may be red for reasons that do not matter; judge the red
ink.

Read the batch file `BATCH` (JSON: a list of {"cand": id, "path": image path}). For each entry, open
the image with the Read tool and classify the RED ink into exactly one class:

- **A**: one or more lines of running body text (sentences, list entries, newspaper column text,
  verse), or a run of at least three whole words of such a line, are red.
- **B**: other text is red: a heading, headline, running head, page number, caption, marginal note,
  table row, advertisement text, title-page text, signature mark, catchword, or fewer than three
  words of body text.
- **C**: the red is not text: a rule, ornament, illustration, border, page edge, fold, stain,
  bleed-through or show-through from the other side of the leaf, noise, blank.
- **D**: the red is only fragments of letters whose words are otherwise blue (a descender, an
  accent, the end of a word whose box stops short), not a whole missing word.

If mixed, choose the first class that applies in the order A, B, D, C. For A, also count how many
printed lines have at least three red words (`lines`). Write one JSON object per line to `OUT`:
{"cand": ..., "class": "A|B|C|D", "lines": n, "note": "<under 15 words: what the red is>"}.
Do not skip any entry; if an image will not open, write class "X". Do not edit any other file.