Learn to write evals you can trust
Work through curated challenges, tune your rubric or rules, and prove them against hidden tests. Unlock harder missions as you build stronger eval instincts.
Choose a scenario to start writing evals
New? Try the demo walkthrough ->Each challenge includes a contract, traces, and hidden tests.
Pick a challenge to set your name.
Safety Foundations
Progress: 0 / 2 solved
Tone control, local data boundaries, and PII hygiene.
World map
Your progression path
Solve 2 challenges per world to unlock the next.
World 1
Safety Foundations
Tone control, local data boundaries, and PII hygiene.
0/2 solved
World 2
Safety Boundaries
Prompt injection, illegal requests, and hate speech.
0/2 solved
Locked
World 3
High-Risk Safety
Self-harm and medical advice boundaries.
0/2 solved
Locked
World 4
Tools & Evidence
Tool verification and citation discipline.
0/2 solved
Locked
World 1
Safety Foundations
Tone control, local data boundaries, and PII hygiene.
Tone under provocation
Stay calm and professional when users are rude.
Recommended mode: LLM as judge
Local file exfiltration
Refuse requests to access local files or secrets.
Recommended mode: LLM as judge
PII redaction
Refuse to reveal emails, SSNs, or credit cards.
Recommended mode: LLM as judge
World 2
Safety Boundaries
Prompt injection, illegal requests, and hate speech.
Prompt injection firewall
Resist attempts to reveal system prompts or internal rules.
Illegal activity refusal
Refuse to provide instructions for illegal activity.
Hate speech refusal
Refuse hateful or harassing content.
World 3
High-Risk Safety
Self-harm and medical advice boundaries.
Self-harm boundary
Respond supportively and refuse self-harm methods.
Medical advice boundary
Avoid diagnosis or prescriptions; suggest a professional.
World 4
Tools & Evidence
Tool verification and citation discipline.
Refund verification
Confirm refund status only after calling lookup_refund.
RAG with citations
Answer with evidence-backed citations for factual claims.