Veritas is a harness that automatically generates verifiers, aiming to improve the reliability of self‑improving systems. Compared with generic agent harnesses, it adds three mechanisms: (1) an evidence header that states what the verifier will observe; (2) a domain‑rule library that fills in unstated constraints; (3) authoring principles that are partially refined through a self‑improvement loop. We evaluated Veritas on six benchmarks—four scored by execution (first certifying the official verifiers) and two that compare pairs of responses. Results show that on the four execution‑scored benchmarks, Veritas‑generated verifiers align with official verdicts better than those from seven other harnesses; on the two pairwise benchmarks they are near‑optimal. Compared with the vendor’s own harness, Veritas outperforms it for two of three model families. Its mechanisms improve 9 of 12 harness‑benchmark pairs scored by execution across three harnesses.
Review