From a smoke test to an auditable pilot

A small run, with a specific purpose

Before comparing reward strategies, I needed to establish that the training components could work together: construct an environment, train briefly, save a model, reload it and step through a held-out scenario.

The 576-step PPO smoke path addresses that integration question. It does not establish that an agent has learned to navigate well. A short successful execution and a successful policy are different kinds of evidence.

Making the run inspectable

I extended that foundation into a pilot experiment runner. The runner checks scenario collections, records configuration and revision information, accounts for interactions, and retains checkpoints and logs. Those records make it easier to connect a result back to the inputs and implementation that produced it.

The held-out path currently uses the first validation scenario for one post-training episode. It is an integration check, rather than the complete checkpoint-validation protocol.

Where the study stands

At the November 5 source snapshot, five reward calculators are implemented, while the runner still accepts only sparse-reward pilots. The November 4 CI run recorded 457 passed, 2 skipped and 17 warnings at revision 6b8e1ce. That is evidence about the software checks at that revision, not a finding about which reward strategy is best.

The next milestone is to connect the remaining experiment modes and produce comparable, traceable runs. Research conclusions will need the agreed evaluation protocol and retained artifacts, not just a training command that finishes.