# RL environment and benchmark Neurarch's editor doubles as a **verifiable environment for agents that design neural architectures**. The action space is the same structured edit set the product agent uses (195 layer types, 75 blocks, add/wire/modify/scale), and the reward comes from the same verifier that protects human users: architecture scoring, guardrails, and shape inference. Every verdict is programmatic; no LLM judge, no human labeling. ## Why the reward is cheap and dense The dense verifier costs 0.27 ms per verified edit on a 47-layer graph (0.05 ms on a 6-layer one), one laptop core. That is roughly 1.3e7 verified edits per hour with no GPU, against roughly 12 five-minute GPU experiments per hour for train-and-evaluate rewards. Low fidelity by design: it proves legality and graph health, not trained accuracy. The numbers are measured and pinned by a CI test. Does legality tracking matter? In a 264-graph study (two seeds, torch 2.8), all 96 designs the verifier blocked crashed in PyTorch at forward time, and all 80 it passed ran clean. Details and reproduction: . ## The benchmark Design-from-spec: each task gives the policy a natural-language spec and a starting graph; the resulting graph is graded programmatically. 12 curated tasks (CNN, MLP, text encoder, fix-broken-attention, deepen-ResNet, add-normalization, scale-under-budget, autoencoder, two-tower retrieval, DLRM CTR, sequence ranking, semantic-search encoder) plus a generated split with anti-gaming grading. - Public leaderboard of frontier models: - Environment server, GRPO training loop, and Colab: ## Run your own agent The environment ships as an open-source server with an OpenAI-Gym-style step/reset interface; any policy that can emit JSON actions can play. Start from the env page above, or open an issue on the benchmark repo: .