Project setup
Repository, tooling and website scaffolded before any training code.
The helper tooling in rl/tools is in place: RunLogger writes one folder per
run with config.json and a metrics.jsonl line per episode, export_policy
writes the trained weights to policy.json, and sync.py copies runs into the
site and builds the index the pages read. A unit test exports a small network
and checks that replaying its JSON by hand gives the same action probabilities
as PyTorch, which is what lets the browser demo be trusted.
The run shown below, _synthetic-example, is fake. It exists so the charts and
the run table could be built and reviewed before the first real run finished. It
has no exported policy, so the demo page falls back to the random agent.
Next up is the REINFORCE implementation itself.