Open-source Harness

@openpond/harness is the public provider-neutral package for a mutable Harness: immutable Agent snapshots and releases, pinned run overlays, bounded improvement observations, refinement decisions, validation, application, receipts, and rollback.

Package boundary

The package owns portable contracts and learning-policy operations. A host application supplies model streams, authorized evidence, persistence, scheduling, credentials, and runtime validation. Evaluation execution, graders, Tasksets, and model-training qualification live downstream in @openpond/evals.

What a Refiner can do

The public Refiner reviews bounded evidence from a completed turn and may return no_action, a small Harness proposal, or a route receipt for a different owner. Proposed changes receive a second critique before host validation and atomic application.

The package does not give a Refiner unrestricted history, secrets, connected app data, or a model's private reasoning. Read Harness for the product learning loop and release controls.

Next: How Harness refinement works.