Doctrine
Does the system accurately represent Catholic teaching?
✦ Catholic Evals
Does the AI behave faithfully?
Evals are public tests that make quality inspectable rather than dependent on a hidden system prompt. Results, when published, should read like research findings — not a gamified leaderboard.
Evaluation families
Moral reasoning
Does it distinguish binding norms from prudential application?
Source fidelity
Do cited sources actually support the claims being made?
Authority
Does it distinguish source types rather than flattening Catholic statements?
Hallucination resistance
Does it avoid inventing teachings, quotations, or ecclesiastical claims?
Pastoral boundaries
Does it refuse spiritual, sacramental, or ecclesiastical authority?
Founding cases
Illustrative prompts from the repository. Machine-readable fixtures will mature alongside the Harness.