Back to Registry
rubric.summary.faithful
SOFT
Scores summary faithfulness to source text via a 3-component LLM rubric: factual accuracy, key points coverage, no hallucinations.
Scorecard
| Determinism | stochastic-judge |
| Evidence Quality | judge-opinion |
| Intended Use | eval-and-train |
| Gating Required | No |
| Permissions | net:llm |
Attack Surface
| injection risk | medium |
| format gaming risk | medium |
| tool spoofing risk | low |
Test Fixtures
9 total
| Type | Count |
|---|---|
| Positive | 3 |
| Negative | 3 |
| Adversarial | 3 |
Metadata
| Version | 0.1.0 |
| Domain | nlp |
| Task Type | summary_faithfulness |
| Contributor | vr.dev |
| Source | tobysimonds.com/research/2025/09/29/Proofs.html |
Use in SDK
# CLI
vr verify --verifier rubric.summary.faithful --ground-truth '{"order_id": "ORD-42"}'
# Python
from vrdev import verify
result = verify("rubric.summary.faithful", ground_truth={"order_id": "ORD-42"})
# API
curl -X POST https://api.vr.dev/v1/verify \
-H "X-API-Key: vr_live_..." \
-d '{"verifier": "rubric.summary.faithful", "ground_truth": {"order_id": "ORD-42"}}'