Benchmark dossier
ByteVerity is evaluated as an evidence system. These numbers separate submitted draft rows from accepted upstream leaderboard rows, and they keep the cost column visible because cost is a central product differentiator for agent governance at scale.
Last checked July 7, 2026 (UTC)
$0.0065 average agent-side cost.
$0.0462 average agent-side cost.
2,091-row official-shaped BM28/V58 package.
Filed across the US, India, and EU.
tau2 agent benchmarks
The ByteVerity rows below are public draft PRs. Accepted-reference rows come from the upstream public leaderboard files. Costs are average agent-side USD per trajectory where the submission reports them.
Custom neuro-symbolic text submission.
ByteVerity submitted
ByteVerity Validity-Synthesa
Accepted upstream leader
GLM-5.2
1.0 point higher pass^1 and about 9.2x lower agent-side cost than the accepted upstream leader.
Draft custom submission; not yet merged into the accepted public leaderboard.
Omission-aware run while 17 task issues were under maintainer review.
ByteVerity submitted
ByteVerity Validity-Synthesa
Accepted upstream leader
GPT-5.5
About 1.76x pass^1 and about 43x lower agent-side cost than the accepted upstream leader.
Omission-aware draft on 80 of 97 banking tasks; not a full standard accepted row.
Accepted public reference
This is not the full leaderboard. It is the subset needed to make the cost and pass-rate comparisons above auditable on the page.
ContractNLI
ContractNLI is document-level natural language inference for contracts. ByteVerity now has an official-shaped BM28/V58 evidence package covering the full 123-document, 17-hypothesis test shape with zero fallback annotations.
Gold labels are used only for offline audit after the prediction file is fixed; BM7/oracle rows remain diagnostic and are not reported as fair submissions.
Evidence package
official_evidence_submission_bm28_predicate_full_v58_20260706. Internal official-shaped BM28/V58 evidence package, generated July 6, 2026.
Sources
The tau2 comparisons link to public GitHub artifacts. ContractNLI performance numbers are from the local ByteVerity evidence package; the public link below is the benchmark dataset definition.