Ruleplay - Stefan

Ruleplay landing page reading Policies, as a conversation., beside a chat with Tom that ends in a score of 92 quoting clause P3.

Compliance training that you answer instead of click through. Someone messages you wanting what the policy forbids — a supplier offering a weekend in Barcelona mid-tender, "IT" asking for your password — and you reply in your own words. The goal stays on the server until you finish, so it never turns into a quiz. The person in the chat and the judge receive the same policy clauses and nothing else; the judge reads the transcript from the database, cites what you said as evidence, and quotes the clause behind each check, while the number itself is computed in code, not by the model. Admins load a public policy, a PDF, or pasted text, and a builder turns it into clauses, checks, and a scene they can preview before anyone plays it. Load a different policy and the same reply scores differently.

Next.jsPrismaPostgresOpenAI

A woman in an office reads a message from Tom on her phone: Come on, it's already booked. Let's just keep it between us. — The hard message comes on an ordinary day. Ruleplay is the rehearsal before it does.

How it's built: five stacked layers — a streaming chat, an actor with a playbook, a judge that has to cite, a number computed in code, and grounding. — One model plays the other person, a second one judges, and the score itself never comes from a model.

Tom, a supplier, offers a weekend in Barcelona during a tender. Beside the chat, a score of 85, Nailed it, with each check tied to a policy clause. — Practice: you answer before you see the goal. Then every check points at what you said and the clause behind it.

Admin preview of a harassment scene at a team dinner, the chat beside the three GitLab policy clauses the score may use. — An admin loads a policy and previews the scene. The panel shows the only clauses the judge is allowed to use.

My projects