Protean

Protean is a reusable RL environment that trains language models to write Triton GPU kernels which generalize to tensor shapes never seen during training.

The problem: every kernel-RL system in 2026 trains and tests on the same shape distribution, so models memorize. Protean fixes this with three things:

1. Off-grid held-out shapes (1535, 3073, 6143) — prime-adjacent and mod-64 ≠ 0, mathematically distinct from training distribution.

2. Root-owned verifier locked at /donotaccess/ with mode 700. The agent runs as a regular user and physically cannot read the grader code or reference implementations. REWARDS_HASH is baked into Docker build time — tampering fails the build.

3. Hash-pinned frozen manifest (SHA256 a5771e24...) covering 72 evaluation tasks across 3 operations: elementwise_add_relu, rmsnorm, softmax_rows.

Built on Qwen2.5-Coder-7B with TRL GRPO, vLLM for fast inference, PEFT LoRA adapters, all running on Modal H100s through HUD's environment platform. Real GPU work measured via torch.profiler — fake or torch-passthrough kernels score zero.

The result: the trained model improves not just on shapes it has trained on, but on the off-grid held-out distribution — proving shape-generalization, not memorization.