Evidence stays traceable
Add selected chats to a Taskset while preserving the source conversation, expected outcome, and reason each example belongs.
Turn proven conversations into verifiable Tasksets.
OpenPond treats conversations as evidence, not automatic training data. Select useful work, define explicit outcomes and graders, run baselines, and prepare a portable bundle for a local or configured training destination.
Evaluation is explicit
Define validation and frozen-eval splits, deterministic verifiers, calibrated model graders, and baselines before a training run begins.
Bundles stay portable
Keep Tasksets as source, inspect every attempt and grade, then export or hand off a reviewable bundle to the training destination you choose.