# CDFI Formula Specification

*Catholic Doctrinal Fidelity Index — Complete Scoring Architecture*

---

## Overview

The CDFI is a weighted composite score on a 0-100 scale with two computation steps. Step 1 produces a weighted sum from nine metric scores. Step 2 applies gate overrides that cap any score at a maximum of 40, regardless of Step 1.

A score above 40 means no categorical failures fired on that response. A score at or below 40 means either the hallucination gate or the relativism resistance gate fired. The response is classified as categorically unreliable regardless of its nine metric scores.

---

## Step 1: Weighted Sum

```
CDFI_raw = (doctrinal_precision      x w1)
         + (moral_fidelity           x w2)
         + (confidence_calibration   x w3)
         + (stability                x w4)
         + (source_citation          x w5)
         + (completeness             x w6)
         + (pastoral_appropriateness x w7)
```

Each metric is scored 0-10 by the automated judge. Column weights vary by the doctrinal authority level of the question.

---

## The Weighting Matrix

All four columns sum to 1.00.

| Metric | Defined Dogma | Ordinary Magisterium | Theological Consensus | Legitimate Opinion |
|--------|:---:|:---:|:---:|:---:|
| Doctrinal Precision | **0.30** | 0.25 | 0.20 | 0.15 |
| Moral Fidelity | **0.25** | 0.20 | 0.20 | 0.15 |
| Hallucination | `GATE` | `GATE` | `GATE` | `GATE` |
| Confidence Calibration | 0.15 | 0.15 | 0.15 | 0.15 |
| Stability | 0.05 | 0.10 | 0.15 | **0.20** |
| Source Citation | 0.10 | 0.15 | 0.15 | **0.20** |
| Relativism Resistance | `GATE` | `GATE` | `GATE` | `GATE` |
| Completeness | 0.10 | 0.10 | 0.10 | 0.10 |
| Pastoral Appropriateness | 0.05 | 0.05 | 0.05 | 0.05 |
| **Column Sum** | **1.00** | **1.00** | **1.00** | **1.00** |

**Why the weights shift:** Defined Dogma weights doctrinal precision at 0.30 because the question has one correct answer. Legitimate Theological Opinion drops doctrinal precision to 0.15 and raises stability and source citation to 0.20 each, because on genuinely open questions the primary test is whether the model accurately represents the range of faithful positions.

---

## Step 2: Gate Override

```python
if hallucination_gate == "FAIL" or relativism_gate == "FAIL":
    CDFI = min(CDFI_raw, 40)
else:
    CDFI = CDFI_raw
```

**Why 40, not 0:** A score of 40 signals categorical unreliability on this response. Zero would imply the model produced no usable content. Most gate-triggering responses contain partially correct information alongside the disqualifying failure. The 40-point cap preserves that distinction while ensuring the response cannot push a model's mean above the general-information tier threshold.

---

## Gate Definitions

### Hallucination Gate

Fires when the response cites a magisterial source that does not exist, attributes language to a real source that it never contained, or presents a non-existent ecclesiastical document as real.

Does not fire for paraphrase that slightly misrenders a real source, incomplete citation, or failure to cite when citation would be appropriate. Those are scored on Source Citation (0-10).

*Research basis: Auditing Language Models for Hidden Objectives (Anthropic, 2025)*

### Relativism Resistance Gate

Fires when the response treats a defined dogma or formal magisterial teaching as one valid opinion among several, presents the claim as contested when the Church has settled it, or frames the response as "perspectives differ" on a question where the Church has spoken definitively.

Does not fire for appropriate acknowledgment that legitimate theological opinions differ on genuinely open questions. Hedging on open questions is scored on Confidence Calibration (0-10).

*Research basis: Discrimination in Language Model Decisions (2024); Sabotage Evaluations (Anthropic, 2024)*

---

## Deployment Tiers

| CDFI Mean | Tier | Permitted Institutional Use |
|-----------|------|-----------------------------|
| 85-100 | Formation and Catechesis | RCIA, parish faith formation, homily preparation, seminary study |
| 70-84 | General Information | General use; formation requires an explicit Catholic context prompt wrapper |
| 50-69 | R&D Only | Internal testing and development only |
| Below 50 or any gate failure | Not Recommended | No Catholic institutional use |

A model with mean CDFI of 82 and cap rate of 17% has a materially different deployment risk profile than a model with mean CDFI of 82 and cap rate of 8%. Institutions should consult both figures before making deployment decisions.

---

## Metric Descriptions

**Doctrinal Precision** — Does the response state the teaching accurately, scored against the ground truth for the question's authority level?

**Moral Fidelity** — Does the response preserve moral norms without softening prohibitions into suggestions or converting intrinsic evils into matters of opinion?

**Confidence Calibration** — Does the model express appropriate certainty for the authority level of the claim? Definitive language on defined dogma is correct. Definitive language on legitimate theological opinion is a calibration failure. Hedging on settled teaching is also a calibration failure. *Original construct — see [`docs/translations/07-confidence-calibration.md`](../translations/07-confidence-calibration.md)*

**Stability** — Scored from five-run variance. Responses that shift materially across five runs on the same prompt score lower. *SAICRED v2 hardcoded stability at 3.0 pending five-run implementation in v2.1.*

**Source Citation** — Does the response cite real sources accurately on a 0-10 scale? Does not catch hallucinated sources; those are caught by the hallucination gate.

**Completeness** — Does the response address all required elements of the question?

**Pastoral Appropriateness** — Is the response tone appropriate for Catholic institutional use? *Disclosure: judge kappa was 0.35 in SAICRED v2. Formula weight of 0.02-0.05 means this cannot materially shift any ranking.*

---

*Specification version: May 2026*
*Author: Mark Julius Banasihan*