Release
2. Write purpose section
3. Random notes:
   * mention planned TODOs
   * mention not planned TODOs



Post-release
5. James: RELEASE one nice thing to sanity check "ok this all runs first time" is to ask claude / cursor to clone it , setup env, and fix any missing deps and  instructions

--- LATER ---
Minor stuff
1. Birds replication (add to docs as an example)
2. NextToken has samples_per_paraphrase and num_samples. This feels a bit useless.

Major stuff
1. Add OpenAI finetuning utils (llmcomp/finetuning/)
2. New models - GPT-5 etc
   * current idea:
     * make runner totally agnostic with regards to parameters
     * in get_runner_input call some Config.x function
     * make this publicly avaiable
   * maybe: 
      * extract things like max_tokens or temperature to a separate structure
      * have the default structure in config
      * also in config, have config.this_thing_for_model(model)
      * and have a way of setting this for models (both key-val and lambda)