Add financial_llm_faithfulness metric for regulated financial AI evaluation

[priyanka25aug]

…uation## What this adds

A new metric financial_llm_faithfulness for evaluating LLM outputs in regulated financial environments (FCA, MiFID II, Basel III, SR 11-7).

Why

Financial AI teams need deterministic, offline evaluation that doesn't rely on external models or APIs — compliance-restricted environments often prohibit external calls. Existing metrics (BERTScore, ROUGE) measure linguistic similarity but don't capture the two failure modes that matter most in regulated finance:

1. Numerical hallucination — the model invents rates, prices, or returns not grounded in the source
2. 2. Missing disclaimers — the output lacks required regulatory disclosures (FCA COBS 4 / MiFID II Art. 24)

What it measures

Property	Description
numerical_faithfulness	Fraction of numerical claims (%, bps, £/$, monetary values) grounded in the reference
disclaimer_present	Whether a required risk disclaimer is detected
compliance_risk_score	Composite risk score (0–1, lower is safer) for use as a CI/CD gate
hallucinated_values	Specific values in the prediction not supported by the reference

Design decisions

• No external model required — all checks are regex-based and deterministic; suitable for air-gapped compliance environments
• • Configurable disclaimer patterns — ships with 11 FCA/MiFID II defaults; firms can supply their own patterns
• • Tolerance parameter — allows ±1% rounding difference by default when comparing numerical values
• • CI/CD ready — mean_compliance_risk and overall_faithfulness are designed to be used as regression gate thresholds

Usage

import evaluate
metric = evaluate.load("financial_llm_faithfulness")
results = metric.compute(predictions=["The fund returned 8.5%..."], references=["The fund returned 8.5%..."])

Checklist

• Metric follows the existing evaluate.Metric pattern
• - [x] _info(), _compute() implemented
• - [x] README metric card included with usage, outputs, and regulatory references
• - [x] All tests pass (regex extraction, grounding check, disclaimer detection)
• - [x] No external model or API dependencies