todo.md

Priority Queue: - batching inferences - A separate thread that batches inferences on a 1-10ms timeout - State implements Hash and we cache the result of inferences - 128-1024 tasks, each running its own game, on a threadpool - Use Actix, InferenceBatcher is an Actor, spawn 1024 "play a game" tasks on a threadpool - too much Dirichlet? - Game API v2: - separate Expert from Game - Automatically implement GameExpert for game and make generate_games work for any game pub struct Hypotheses { pub actions: Vec, pub priors: Vec, pub q: f32, } struct TrainingExample<S, A> { state: S, hypotheses: Hypotheses q: f32 }

    trait Game<S, A> {
        symmetries(&TrainingExample<S>) -> Vec<TrainingExample<S>>
    }

    State and Action implement Serialize, Deserialize to the format used in training

    - 
- Fill in q values at the end of the game
- Emit all symmmetries of a game as examples



- lift as much logic as possible into seraphim core
    - Librarize the Python half that all lives in train.py atm
- tfrecord proto gencode should be in core
- move TTT to separate crate
- benchmarks
    - profile performance
    - triage performance

- investigate j-curve in readouts:
    - 100 has more draws than 600
    - it starts going up again around 2000, nearing 100% draws by 10,0000
    
- loss = mse + cross_entropy + l2 regularization
    - L2 regularization 10e-4

- hardware accelerate inference
    - debug slow inference perf on gpu
    - batching?
- summary metrics & checkpointing for Tensorboard   
    - accuracy metric

- unit tests    
    - Do the posterior possibilities add to one?
    - testing framework for implementors
- auto tournament for model comparison
- multithread search
    - replace petgraph with a custom lockless search tree

- debug mode:
    - allow initialization from any state saved in a file
    - better (user-defined) parsing of next action in interactive games

Large Features: contrib - Contribute TFReader

interactive:
    - flesh out debugging & interactive game traits
    - history mode/undo/branch exploration
    - implement an example GUI

search:
    - search benchmark
    - discard untaken edges in the search tree when advancing down a node
    - Multithread search

python side:
    - many implicit contracts between the Rust and Python halves. Document them, librarize them
    - performance and strength benchmarking tools