Verifier-backed workflow data for professional software
Professional software workflow data needs verifier audits, pass@k distributions, failure traces, and contamination controls to be useful.
Professional software workflow data needs verifier audits, pass@k distributions, failure traces, and contamination controls to be useful.
For agents working in professional software, the bottleneck is not only RL algorithms. It is the supply of verifiable workflow environments.
Good professional software workflow data has verifiers, eval splits, failure traces, and evidence that it improves agents.