Describe your AI/agent system and get one prioritized PASS / WARN / FAIL report — reward-hacking, model-collapse, agent-herding, fake A/B lifts, RAG rot, biased causal claims — each with the fix. Every check is measured, not assumed. Run it free below, in your browser.
Fill in whatever parts of your system you have (pre-filled with a failing example). It runs entirely in your browser — nothing is sent anywhere.
Answer the ones you can — skip the rest. The defaults are a deliberately broken example so you see a full report. Hover units for help.
Your A/B test — conversions out of users, per variant.
Model-collapse risk from synthetic / self-generated data.
Multi-agent / ensemble herding (skip if single-agent).
Reward-hacking / Goodhart — when a metric becomes a target.
Causal/attribution — one variable you "control for". Adjusting for the wrong kind injects bias.
Each is a proven, measured tool — and we run all of them on ourselves, publicly. See our own self-audit →
Is a reported lift real, or noise?
Is the model collapsing on itself?
Will your agents herd?
Is the metric/reward gamed?
Is the causal number identified?
Is your RAG store rotting?
Agent-memory health.
When to quit a depleting effort.
pip install