Runmark

Inspiration

Reward curves are persuasive even when they are wrong. A run can keep improving its training score while generalization plateaus, value estimates drift, or the verifier becomes exploitable. Runmark turns that gap into a clear decision and a short evidence trail.

What it does

Runmark is a visual audit desk for reinforcement learning runs. It compares training reward with held-out accuracy, zero-variance fraction, and gradient usability. The demo includes three deterministic scenarios: healthy learning, saturation, and reward hacking. Builders can switch chart windows, mute individual signals, open evidence checks, create a named audit, and move into a notes view.

How I built it

I built Runmark as a React and Vite prototype with a local fixture layer so the story stays reproducible without a backend. The charts are lightweight SVG and the UI uses CSS and Phosphor icons. The layout is responsive and includes dark and light themes.

Challenges I ran into

The hard part was making a dense training dashboard readable at a glance. I used one accent color, a strong information hierarchy, and plain-language diagnosis cards so the interface communicates a decision instead of just showing more metrics.

What I learned

Reward is only useful while it agrees with evidence that generalizes. A small set of aligned signals can be more actionable than a larger wall of telemetry.

What's next

Connect the audit board to real checkpoint streams, add verifier replay and checkpoint comparisons, and persist the evidence trail for team review.

Built With

• css
• javascript
• react
• svg
• vite