Glossary

Term	Meaning
Acting agent / actor	The system attempting tasks; implemented by your pipeline
Adapter	A directory of context, skills, roles, or tools, with a manifest defining its training contract
Template / seed	The starting adapter structure and its fixed/learnable rules; a seed can already contain useful behavior
Host	An agent application or custom loop consuming an adapter
Pipeline	Your run(task, adapter) function that executes one task
Judge	Your scorer producing a numeric score and learner-visible feedback
Evolver / learner	The method that uses training evidence to revise an adapter
Trainer	The outer execution loop, checks, accounting, history, and recovery
Learning material	Recorded task inputs, outputs, judge results, and available trajectories supplied to an Evolver
Trajectory	Messages, tool interactions, or logs from an acting attempt, when the pipeline supplies them
Step / batch	One main rollout batch followed by one Evolver update
Epoch	One pass through the configured training split
Specification check	Validation against the template’s file rules and optional checker; not a performance test
Candidate	A proposed adapter version; a method may test multiple candidates inside one step
Adapter tree	Git history of versions belonging to a template
Checkpoint	A tagged version saved during training
Final	The adapter selected at finalization; not universally the highest-scoring checkpoint
Learner state	Method-owned files persisted across steps and saved for recovery
Session resume	Continuing a learner host session; separate from crash recovery
Deployment	Writing adapter exports into a host workspace
Conversion	Emitting a native plugin package for a target host
MWM block	An ordered group of environment episodes between learner reviews
Evidence store	Task-owned episode records consumed by MWM under its native protocol