tcpsyn c7aac1b373 Make AI callers more dynamic with arcs, seeds, style gradients, and learning ...

- Conversation arc system: callers evolve through Opening → Building → Turn →
  Depth → Wind down stages, with style-based timing offsets
- Problem seeds: 80 short concept strings that the LLM expands into unique
  stories each time (50% of PROBLEMS picks use seeds vs full entries)
- Style gradients: 4 composable axes (energy, honesty, humor, confidence)
  with 625 combinations, used 50% of the time alongside handcrafted styles
- Call duration tracking: records metrics to data/call_metrics.json, learns
  which pools and style axes produce longer calls, nudges future picks ±15%

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

2026-03-12 06:21:09 -06:00

419 KiB

Raw Blame History

View Raw