# Faultprint AI Depth Mandate

Date: 2026-07-05

## Purpose

The operating library and postmortem library must both become deep enough on AI-specific institutional conditions to support a distinct AI-era product line. General automation, scoring, and rule-based workflow evidence is valuable, but it is not enough. AI-related conditions must be deliberately coded as their own analytical layer.

## Required Distinction

Faultprint must separate three categories:

1. Rule-based automation
   - Deterministic eligibility rules, workflow gates, forms, policy criteria, and fixed scoring logic.

2. Algorithmic/scoring systems
   - Risk scores, telematics scores, screening scores, fraud scores, ranking systems, identity matching, and composite classifications.

3. AI / model-mediated systems
   - Machine-learning models, generative AI, AI-assisted review, AI detectors, biometric/liveness systems, predictive models, automated decision systems, model governance, model monitoring, fairness testing, and human-in-the-loop controls.

Rows may overlap, but the AI layer must not be diluted by broad automation language.

## Operating Library AI Target

The operating library should include a dedicated AI operating tranche with at least 40-60 additional AI-specific condition rows across domains such as:

- HR and hiring AI
- Insurance AI and telematics models
- Healthcare utilization review AI
- Fraud/risk scoring AI
- Identity verification and biometric AI
- Education AI detection and proctoring
- Content moderation AI
- Credit/underwriting AI
- Public-benefits AI or automated decision support
- Customer-service and generative AI systems

Each AI operating extraction must identify:

- AI function or model role
- Decision proximity
- Human review claim
- Contestability/correction path
- Model monitoring or audit claim
- Fairness/bias control claim
- Evidence gap
- Falsification test
- Boundary

## Postmortem Library AI Target

The postmortem/failure library should include a dedicated AI failure tranche with known public cases where AI, model-mediated classification, automated decisioning, or algorithmic systems materially contributed to institutional harm or breakdown.

Candidate postmortem domains include:

- Automated welfare or benefits decisions
- AI hiring discrimination or screening
- Biometric/face recognition misidentification
- Healthcare AI/utilization management disputes
- Education AI/proctoring/detection controversies
- Public-sector risk scoring
- Credit/insurance model discrimination
- Content moderation or platform enforcement failures
- Fraud detection false positives

Each postmortem AI extraction must identify:

- Model/system function
- Failure mechanism
- Harm pathway
- Contestability failure
- Evidence source
- Institution response
- Comparable operating conditions

## Product Relevance

The AI product line should not say: "This institution uses AI, therefore it is risky."

It should say:

"We identify where AI or model-mediated systems have become decision gates, then test whether the institution has evidence of correction, monitoring, fairness, review quality, and human accountability."

## Immediate Work Rule

Until the AI layer is deep enough, future extraction work should include AI-focused batches rather than only broad cross-sector expansion.