Trust Metrics
What is measured
Section titled “What is measured”Generated proposal text is decomposed into atomic claims. Clause citations and numbers are verified deterministically against the FAR/DFARS catalog and the cited sources. Remaining claims go through a batched source-fidelity check. Unsupported claims are flagged or rewritten; sections with unsupported claims escalate to human review.
What is published
Section titled “What is published”The public trust page publishes real, measured numbers only: citation resolution rate, claim support ratio, sections reviewed, escalation rate, and the results of a seeded adversarial eval (seven failure modes mapped to real, court-documented AI-hallucination cases). Unverified claims are excluded from both numerator and denominator. When no proposals have been reviewed, the page publishes zeros and says so.
The live numbers are at govconhouse.com/trust.
What is not claimed
Section titled “What is not claimed”- No zero-hallucination claim. No honest system can.
- The NLI (natural-language inference) eval lane is a paid lane; the page states its status instead of implying it always runs.