RL ENVIRONMENTS FOR FRONTIER CODE MODELSRL Envs for Recursive Self-ImprovementEXPERT ENGINEERING DATA FOR FRONTIER CODE MODELSRL ENVIRONMENTS FOR RSI-CRITICAL CAPABILITIESEXECUTABLE TRAINING SIGNAL FOR FRONTIER CODE MODELS

Closing the data gap for RSI.

Horizon partners with experts in low-level coding infrastructure domains on the critical path to increasing the rate of recursive self-improvement.Kernel, compiler, inference, and systems engineers produce the rare training signal frontier models need to improve the infrastructure beneath their own intelligence.The next gains in model capability depend on systems work that never appears cleanly in public code. Horizon turns that expert work into executable learning environments.Frontier models cannot automate improvement until they master the low-level capabilities that constrain it. Horizon sources those capabilities from the engineers who build them.From pre-training algorithms and CUDA kernels to inference engines and agent harnesses, Horizon captures expert work across the entire recursive self-improvement loop.Public repositories show finished code, not the decisions that make difficult infrastructure work succeed. Horizon captures those trajectories as verifiable training data.We build the low-level coding environments frontier models need to take on more of the research and engineering required to improve themselves.By focusing expert data production on RSI-critical infrastructure, Horizon increases the rate at which frontier models acquire their next generation of capabilities.Rare systems expertise becomes useful training data only when it is executable, inspectable, and rewardable. Horizon makes that transformation possible.Horizon concentrates expert trajectories, RL environments, and verifier design on the engineering hotpath for recursive self-improvement.