Show HN: An open-source, RL-native observability framework we've been missing (github.com)

An open-source observability tool called verifiers-monitor (installable via pip) was announced for real-time monitoring of reinforcement learning training and evaluation. It plugs directly into the verifiers ecosystem with a one-line wrapper around environments, spins up a local dashboard (http://localhost:8080) and streams WebSocket updates so you can see progress, stalls, and live reward charts as rollouts complete. The UI and API surface show per-example pass/fail status, full prompts and completions, reward breakdowns and attribution, multi-rollout comparisons, and session-to-session metric tracking. Technically, the package exposes a MonitorData API that lets you programmatically fetch sessions, rank worst-performing examples, detect instability by reward variance, inspect best/worst rollouts and tool calls, and export results to pandas for custom analysis. That combination of real-time visibility, per-example traceability, and easy data export addresses a recurring pain point in RL experiments—poor observability—making debugging faster, improving evaluation fidelity, and helping teams identify high-variance or reward-misaligned cases earlier. For researchers and engineers building verifiers-based environments, this toolkit promises tighter iteration loops, clearer failure modes, and a practical path toward standardized, reproducible RL evaluation workflows.