video-digest-segment

LLM stage — no scripts. Input: <item_dir>/transcript/timeline_30s.txt (from ingest). Output: <item_dir>/talks.json.

This is the fragile stage — read this before running

A single pass over a braided feed (stream cutting between stages, re-shows, MC filler) drops real talks. Always run the verification pass below.

If an authoritative schedule exists (conference program page, description chapters, chapters.txt), save it to <item_dir>/schedule.txt — titles/speakers/times in order. It's reconciliation ground truth. No schedule → derive talks purely from content.
Run the segmentation prompt: references/segment-prompt.md (fill {ITEM_DIR}, {CONTEXT}, paste schedule facts). One agent for the whole stream.
VERIFY (mandatory): walk the schedule (or the talk list you produced) in order; for EACH expected talk, grep transcript/transcript.tsv for the speaker's self-intro or a distinctive title phrase and confirm a matching entry exists in talks.json. A scheduled talk braided with cutaways must still appear (kind main/talk) even if windows overlap. Timestamps drift — trust transcript CONTENT over scheduled clock times.
Sanity-check output: talks ordered by start; start/end "HH:MM:SS"; leading applause trimmed; kinds valid.