Reinforcement learning for IBM Z mainframe

Run the existing code.
Pay less for it.

A Gymnasium-compatible RL environment that finds MIPS optimisations in COBOL batch programs — using real z/Architecture execution, not estimates.

Book a call → View on GitHub ↗
refinery — episode 1,240 · HERCULES pool · MVS 3.8j
average MIPS
reduction per program
monthly saving on a
500-MIPS workload
functional correctness
architecturally enforced

IBM Z manages $3 trillion in daily commerce. Enterprises running it are billed by the MIPS — every inefficient loop, every redundant compute, every unoptimised sort adds directly to the monthly bill. The engineers who know how to fix it are retiring and not being replaced. Refinery closes the loop: an RL agent that runs your COBOL against a real z/Architecture emulator, reads real SMF telemetry, and discovers optimisations no static analysis tool can find.