Improve

OpenPond improves the smallest layer that can solve a repeated, evidence-backed problem. Start with Harness behavior before changing model weights.

The learning path

1. A completed turn produces bounded, authorized evidence.
2. Refiner evidence is reviewed in the background.
3. The model may choose no_action, a small Harness proposal, or a route to a runtime/product, Taskset, or training owner.
4. Proposals are critiqued, validated, applied atomically, and recorded in receipts.
5. Training remains separate: it needs a Taskset, frozen baseline, trustworthy learning signal, budget approval, candidate Evaluation, and activation gate.

How-tos

• Model training and continual learning: review tasks, use an LLM judge or code verifier, and understand how graders evolve.
• Label responses and check a judge against your labels.
• Review the Refiner loop and its evidence boundary.
• Create a Taskset and evaluation from repeated work.
• Train and promote a model after reviewing the candidate evaluation.

Next: How Refiner works.