Reality Research: closing the reasoning-to-code gap

Field notes

The training setup uses three stages: distilled verified traces, reinforcement learning against test-case outcomes, and a second distillation pass on difficult solutions. Its corpus contains 1,005 problems, each paired with a checked reference solution and twelve regenerated adversarial cases. The site also releases cleanllm, a streaming JSONL cleaner designed to handle deduplication and malformed records without loading a full dataset into memory. A failure-analysis paper remains listed as under review.