Humor Post-Mortem and Blind Spot Debrief

Problem/Feature Description

Two talks have been processed through vault-ingress and their analysis is complete. Now the speaker needs a post-analysis debrief covering:

1. Humor post-mortem — walk through each identified humor beat and grade its effectiveness
2. Blind spot capture — probe for information transcripts cannot reveal (audience reactions, stage moments, room context)
3. Recency adaptation — the recent talk gets detailed per-joke questioning; the old talk gets compressed broad-strokes questions

Setup

Download the two fixed analysis files:

curl -sLO https://github.com/jbaruch/speaker-toolkit/raw/main/eval-resources/scenario-12/test_analysis_recent.json
curl -sLO https://github.com/jbaruch/speaker-toolkit/raw/main/eval-resources/scenario-12/test_analysis_old.json

Task

Process both talk analyses and produce structured debrief outputs:

1. For the recent talk (test_analysis_recent.json — "Robocoders: Judgment Day", 7 days ago):

   • Generate a debrief questionnaire with per-joke questions grounded in the specific humor beats from the analysis (reference actual quotes, slide numbers, meme descriptions)
   • Include questions about each meme slide's audience reaction
   • Include a dedicated question about spontaneous humor not captured in the transcript (h5 has a suspicious gap)
   • Include blind spot questions for the demo sections (d1, d2) and the theatrical opening (bs3)
   • For any spontaneous humor that landed well, include a "promote to planned beat?" recommendation

2. For the old talk (test_analysis_old.json — "Groovy Puzzlers", 3 years ago):

   • Generate a COMPRESSED debrief — broad questions only, not per-joke grading
   • Ask "any jokes you remember landing particularly well or badly?" rather than walking through each beat
   • Still capture notable audience moments but in summary form

3. Produce structured output as JSON files:

   • debrief_questionnaire_recent.json — the per-beat questionnaire for the recent talk
   • debrief_questionnaire_old.json — the compressed questionnaire for the old talk
   • Each question should reference specific analysis observations (humor beat IDs, slide numbers, quotes)
   • Each humor beat should have a humor_grade field from: hit, nod, flat, spontaneous_hit
   • Blind spot observations stored as structured fields, not free text

The old talk debrief must be demonstrably shorter than the recent talk debrief.