Collected sources and patterns will appear here. Add from search, explore, or the patterns library.
Environment -> PhaseMetrics
Run the agent in the environment for a fixed number of steps under a 'training' phase (with exploration and parameter updates) followed by an 'evaluation' phase (without exploration or updates) to generate separate diagnostic statistics.
Problem it solves
Mixing training and evaluation metrics masks the agent's true exploitation capabilities and introduces bias in progress tracking.
Consumes
Emits
The real projects this mechanism was found in. Attribution is the point — this is how the best teams actually do it.