Restore Structure to Flattened Meeting Notes

Pass one, one question per adjacent line pair, all in a single request. Tag each line in code with a short id before sending, like L014| at the start of the line.

{
  "model": "jev-latest",
  "state": "{{tagged-lines}}",
  "questions": {
    "L002_continues_L001": {
      "type": "noul",
      "instructions": "Does line L002 pick up mid-sentence, continuing a sentence left unfinished at the end of line L001?"
    }
  }
}

Merge every pair that scores above 0.7 into one block. Pass two, one Choice per merged block, again in one request:

"B01_type": {
  "type": "choice",
  "instructions": "What kind of content is block B01?",
  "criteria": {
    "heading": "Short label or title naming the document or a section",
    "paragraph": "Running prose with complete sentences",
    "list_item": "One entry in a parallel series of items",
    "quote": "Words attributed to a person or source",
    "code": "Computer code or terminal commands",
    "callout": "A warning, tip, or important note that interrupts the flow"
  }
},
"B01_heading_level": {
  "type": "choice",
  "instructions": "If block B01 is a heading, what level is it?",
  "criteria": {"title": "The document title", "section": "A main section", "subsection": "A subsection", "none": "Not a heading"}
},
"B01_is_step": {
  "type": "noul",
  "instructions": "If block B01 is a list item, is it a step in an ordered sequence rather than an unordered bullet?"
}

Render Markdown in code from the answers. Never let the model rewrite content; blank lines and explicit markers like # or 1. that already exist are kept as direct evidence and not reconsidered. Two requests recover headings, lists, quotes and callouts from a pasted transcript of {{meeting-notes-source}} for a fraction of a cent.