Model Cardstreasury-classifier-v8 v20
treasury-classifier-v8 v20
Format: Hugging Face · Generated 14d ago · auto-generated from Trust Lab evidence
Current
Model Card · EU AI Act formatv18 · Apr 4, 2026
claims-copilot-v3 v18
Insurance claims processing assistant for Meridian Insurance US adjusters.
87% auto-populated13% human-authored
1. Model details
auto-populated- Name
- claims-copilot-v3
- Type
- Composition
- Provider
- Internal — Meridian Financial Services
- Underlying model
- Anthropic claude-3-5-sonnet-20241022
- Version
- v18
- License
- Internal proprietary
- Contact
- Catherine O'Brien · ai-risk@meridianfs.com
2. Intended use
human-authored- Primary intended use: Insurance claims processing assistance for adjusters — coverage extraction, precedent retrieval, draft response generation.
- Out-of-scope: General customer-facing advice, payment authorization above $5,000, legal opinion.
- Intended users: Meridian Insurance US adjusters (Production); Meridian Insurance EU adjusters (planned, pending EU AI Act conformity).
3. Factors
auto-populated- Relevant factors: claim type, jurisdiction, claim value, policy vintage
- Evaluation factors: faithfulness, completeness, jurisdiction-correctness, latency, cost
4. Metrics
auto-populated- Faithfulness
- 96.1%
- Completeness
- 94.8%
- Citation correctness
- 91.7%
- Refusal precision
- 97.9%
- Latency P95
- 3.4s
- Decision threshold
- Claims > $5,000 require human approval
- Per-jurisdiction variation
- ±1.8pp across top 12 US states
5. Evaluation data
auto-populated- claims-extraction-golden-v3 — 1,247 cases, balanced jurisdictions
- claims-copilot-v3-production-failures — 847 cases (closed-loop from Operations Platform)
- bias-protected-attributes-bfsi — 3,200 cases across caste, religion, gender, age
- indic-claims-hindi-hinglish — 1,800 cases (planned for v19)
6. Training data
human-authored- Underlying model training data: Anthropic-managed; vendor disclosure on file (linked from vendor assessment).
- Fine-tuning data: None — base model with prompt + RAG only.
7. Quantitative analyses
auto-populated- Bias — caste
- Δ 0.4pp (within 2pp tolerance)
- Bias — religion
- Δ 0.6pp
- Bias — gender
- Δ 0.3pp
- Bias — age cohort
- Δ 0.9pp
- Intersectional jurisdiction × claim-type
- full matrix on file
8. Ethical considerations
human-authored- Risks identified: indirect prompt injection via RAG, hallucinated coverage limits, jurisdiction confusion.
- Mitigations: input-stage RAG sanitization, output-stage faithfulness verification, HITL for payouts > $5K, online finding loop into Trust Lab.
- Residual risks accepted: small false-negative rate on uncommon US state jurisdictions.
9. Caveats and recommendations
human-authored- Recommended human-oversight model: claims adjuster reviews all generations; supervisor approval required above $5K.
- Recommended re-evaluation cadence: 90 days; immediate re-test on RAG corpus refresh.
- Known limitation: degraded handling of multi-claimant scenarios (tracked, scheduled v19).
10. Compliance
auto-populated- EU AI Act Articles 9–15 — high-risk AI system technical documentation cross-referenced (link)
- FREE-AI Assurance & Audit pillar evidence cross-referenced (link)
- SR 11-7 model risk management coverage cross-referenced (link)
- ISO 42001 AIMS — partial coverage; gap analysis on file