Context

DeepEval evaluates LLM outputs on faithfulness, relevance, hallucination, etc. A related dimension that's missing: trust scoring — how trustworthy is the output based on what sources it used?

Two responses can score identically on faithfulness but have very different trust profiles:

• Response A sourced from SEC filings (high trust)
• Response B sourced from unverified blog posts (low trust)

Possible metric

A TrustScoreMetric that evaluates:

• Source tier — were the retrieval sources authoritative (T1-T2) or unverified (T4-T5)?
• Provenance completeness — does the output carry metadata about its origin?
• Verification status — was the output human-reviewed?

from deepeval.metrics import TrustScoreMetric

metric = TrustScoreMetric(
    threshold=0.7,
    source_tiers={"SEC filings": 1, "news": 3, "forums": 4, "AI inference": 5}
)
test_case = LLMTestCase(
    input="What was Q3 revenue?",
    actual_output="Revenue was $4.2B",
    retrieval_context=["SEC 10-Q filing: Revenue $4.2B"]
)
metric.measure(test_case)
# trust_score: 0.95 (T1 source)

Why this matters

1. EU AI Act Article 50 (August 2, 2026) — compliance requires trust transparency
2. Enterprise RAG systems need to differentiate high-trust vs low-trust outputs
3. Extends DeepEval's coverage from quality metrics to trust metrics

Reference

• AKF defines source tiers (T1-T5) and trust computation

Would the team consider trust scoring as an evaluation dimension?