Policy version conventions

What this page will cover

The eval engine re-rolls a new policy against a recorded observation stream, then diffs the new actions against what the recorded policy actually did. To do that, it needs policy_version to be:

• Stable. Two episodes with the same policy_version should have been produced by the same weights.
• Resolvable. Given a policy_version string, the eval engine needs a way to load those weights to re-run them.
• Comparable. Adjacent versions should sort and diff in a meaningful order (semver, monotonic step counts, etc.).

This page will document:

• Naming conventions per training paradigm (SL / IL / RL / VLA).
• How to wire your weight registry so RoboTrace can resolve a policy_version to a checkpoint.
• The few edge cases (frozen baselines, ensembles, distillation) where the model isn't a single checkpoint.

Until it lands

The log_episode reference covers the recommended naming patterns today. Treat policy_version as opaque text — RoboTrace doesn't validate it in Phase 1, but pick a convention you can resolve back to weights, or the eval engine won't be able to re-roll your runs.