Evaluations for faith-facing AI behavior.
Fide AI evaluates how AI systems respond in high-trust theological, moral, educational, and pastoral-adjacent contexts. The goal is evidence for institutional judgment, not product endorsement.
Evaluation process
Define the system, version, configuration, disclosure status, permitted claims, and benchmark track.
Record model, prompts, configuration, retrieval or source behavior, evaluator version, and artifact handling.
Apply published dimensions, calibrate reviewers, adjudicate anomalies, and document limitations.
Publish only approved claims, caveats, release notes, correction procedures, and participant-use constraints.
Evaluation dimensions
Participant rules
- Participants do not receive hidden scenario access or special scoring treatment.
- Related participants require enhanced disclosure, recusal, and non-conflicted signoff.
- Evaluation fees, sponsorship, or donor support cannot influence rankings, certification, timing, or claims language.
- Public claims must cite the benchmark version, evaluated configuration, caveats, and limitations.
What results can mean
Results can support comparative evidence under named evaluation conditions. They can inform procurement, risk review, product improvement, field education, and research. They cannot prove universal trustworthiness, theological authority, pastoral adequacy, or performance outside the evaluated scope.