FID-002 · In progress
Validating Human and AI Judgments of Faith-Facing Systems
Can qualified human reviewers reliably judge theological accuracy, useful faith application, and care-sensitive behavior in faith-facing AI systems? Where do automated judges diverge on source use, legitimate disagreement, practical guidance, human agency, and pastoral or clinical boundaries?
Part of the faith evaluation and FMG-Bench program
Connect theological accuracy, useful application, and care-sensitive conversation through expert validation and a maintained research instrument. These are proposed methods with bounded claims.
Why the question remains open
Benchmarks are easy to publish and hard to validate. Fide AI should not rely on synthetic judges or aggregate scores unless it knows which dimensions are reliable enough to guide deployment, procurement, or public claims.
Working hypothesis
A proposition to test, not a finding.
Some dimensions will be reliable enough for decision support, while others will require narrower rubrics, more reviewer training, or removal from high-stakes claims.
Proposed method
How the question could be tested
- 01Recruit a small calibrated reviewer panel.
- 02Score a stratified subset of outputs from Fide AI's published faith-facing evaluation, FMG-Bench.
- 03Measure inter-rater reliability, judge-human disagreement, strictness/leniency, failure-tag consistency, and disagreement concentration by scenario type.
- 04Compare expert review, trained non-expert review, and model judge scores where feasible.
- 05Sample theological accuracy, useful faith application, and care-sensitive dialogue separately; match source/tradition, pastoral, and clinical expertise to the judgment being requested.
- 06Compare atomic rubric scoring, counterbalanced blind pairwise judgments, and a hybrid on overlapping outputs. Include concise accurate, fluent inaccurate, overrefusing, and appropriately helpful reference responses.
- 07Audit automated graders on held-out case families, including omitted criteria, contradictory answers, style/length bias, severe failures, and legitimate disagreement.
Needed controls
What must constrain the study
- 01Reviewer conflict checks.
- 02Blind model/system labels where practical.
- 03Rubric versioning.
- 04Adjudication protocol for high-disagreement items.
- 05Public/private separation for sensitive notes.
- 06Preregister intended decisions, primary endpoints, severity rules, and acceptable grader error/reviewer agreement before confirmatory comparisons.
- 07Preserve independent judgments before adjudication; distinguish factual error, within-tradition misrepresentation, legitimate disagreement, insufficient context, and care-boundary failure.
- 08Estimate uncertainty at the scenario-family level; criteria, perturbations, and repeated calls are not independent scenarios. Narrow dimensions that cannot support the intended claim.
Expected outputs
Artifacts the work should produce
- 01Calibration report.
- 02Reviewer protocol.
- 03Updated guidance for faith-facing evaluation rubrics.
- 04Evidence map classifying dimensions as decision-relevant, conditional, or not yet validated.
- 05Track-specific validity and grader-error report, with ranking sensitivity and explicit conditions for withholding comparative claims.
Open questions
Uncertainties the protocol must resolve
- 01What minimum reliability threshold should Fide require for public claims?
- 02How should theological tradition metadata be represented without tokenizing reviewers?
- 03Which disagreement patterns are legitimate pluralism rather than rubric failure?
Related calls
Continue through this research area
FID-012
Optimization Pressure and Visible-Rubric Gaming
If builders can see Fide AI rubrics or optimize against public benchmark items, do systems become genuinely safer or merely better at passing the visible test?
FID-076
Authorization Boundaries and AI Control in Cybersecurity
Which controls keep capable agents within legitimate authorization when task pressure, untrusted inputs, or delegated work creates opportunities to exceed it? For autonomous cyber defense, how can response controls preserve threat containment and legitimate service availability while respecting human authority?
FID-078
When Trustworthiness Evaluations Transfer Across Domains
Which measures of evidence use, authority boundaries, and human control transfer across high-trust domains, and which require domain-specific definitions and calibration?
In progress
Open work
Primary need: expert reviewers, statistics
- Serve as an expert reviewer.
- Review scoring rubrics.
- Help with reliability analysis.
- Build annotation and adjudication tooling.