Read the pack-authoring guidance. Same text as the asistonto:// resources, for clients that cannot read resources. Topics: pack-format, four-part-questions, paraphrases, calibration, workflow, labels, example, all.

Assemble one pack question from the four parts, placing each where the retrieval measurements say it belongs: the heading label and the question are indexed, the provenance and the stakes are not. Returns the question object plus what is still missing from it.

Parse a corpus document into pack questions. Accepts both shapes the asistonto corpus uses: the four-part blockquote format and the older ### QC-1 <title> headings with numbered sub-questions. Reports which one it read.

Lint the paraphrases of one question before they go into a pack: count against the measured curve, phrasings that borrow document vocabulary from the provenance, phrasings that fit a sibling question better, duplicates and keywords.

Validate a pack against the schema the live copilot loads, and against what measurement says will under-perform silently. Errors mean it will not load; warnings mean it will load and then quietly do less than it should.

Write a pack to a directory: brief.json, paraphrases.json and optionally background.txt and labels.json. Validates first and refuses to write a pack that would fail to load.

Read a pack from disk the way the live server does, and validate it. Use this to review or extend a pack that already exists.

Render questions back into the four-part markdown a person can read and correct. Keep it next to the pack: the pack is for the machine, this is for the room.

How many questions a pack needs before anything in it can fire, for a given gate. The live gate is a z-score over the pack's own scores, so a small pack has a ceiling below it and stays dark.