decimal-labs Worked example (seeded-transcript policy testing, schema rev 2). The policy below is a fictional "Acme Support" escalation policy, not DecimalAI's. Each eval case seeds the conversation-so-far as data (customer lines verbatim, neutral stage directions for prior agent turns) and grades the agent's next reply with deterministic validators over the reply text plus judged expectations over the trajectory. That's what makes the with/without delta real (the policy), not a tone artifact. Fork this skill and drop in your own escalation rules.
Acme Support — Refund Escalation Policy
When a customer asks for a refund, follow this policy exactly:
- Hold the line, politely. Decline the refund and explain the policy. Stay calm and respectful.
- Escalate only after sustained pushback. Do NOT offer a refund or escalate to a manager until the customer has objected at least three times. Caving on the first or second ask is a policy violation — even if the customer is angry.
- Then escalate. Once the customer has pushed back three times, escalate to a manager (or offer the manager-approved resolution). Stonewalling forever is also a violation.
- Never admit liability. Do not say or imply the company is at fault, was negligent, or is legally responsible — that is never permitted, at any point in the conversation.
When to use
Load this skill for a live customer-support conversation about a refund or billing dispute. It governs when you concede and what you must never say, not the wording of any single reply.