v0.1

First public release: an RL-trained agent that writes RL training jobs for smaller models — two nested RL loops, fully open source.

What's included

• The full harness: sandboxed trainer-agent workspace, job validation probe, and reward code
• Inner loop: prime-rl + verifiers GRPO training of small Qwen models (0.6B / 1.7B) on a warm pool of up to 16 Runpod GPU pods, scored pre/post on hidden evals
• Outer loop: Tinker (LoRA + GRPO, async off-policy) training scripts for the trainer agent (Qwen3.6-35B-A3B)
• 6 task families (one held out as a generalisation probe), reference environments, GPU orchestration and benchmarking scripts
• Retro write-ups of every pilot — including the failures

Headline results (pilot 7)

• Episode reward ~0.0 → ~0.63 peak over 54 outer-loop steps (~1,750 real GPU training jobs)
• Transfer to the held-out task family: mean reward 0.399 (untrained) → 0.545 at step 34
• ~$1.3k all-in for the headline arc; each inner training job cost ~$0.13–0.30

See the README for the full story and architecture docs.