Skip to content

Repository files navigation

Verdict Bench

Verdict Bench is a public research harness for certificate-bearing routing and evidence refusal in tool-using language-model agents. It applies a posterior-free, anytime-valid control stack to the public Tau benchmark and records enough evidence to audit every candidate decision.

Current release status: implementation and preregistration only. The included results are deterministic [SIM] fixtures for CI and interface testing. They are not measured benchmark performance.

What It Tests

Verdict Bench compares six declared agent-control arms:

  • GREEDY
  • BON-2
  • BON-4
  • EPS-0.35
  • RAW
  • CERT

CERT-EI is an optional comparator. The superseded heuristic is retained only as the provenance label HEURISTIC-v0; it cannot execute and is excluded from public headline tables.

The certified arm logs pairwise accept/refute evidence, fixed-scale W trials, component-ledger evidence, e-BH retirement thresholds, T8 context status, and every evidence refusal. Proxy scoring uses visible policy, conversation, and tool state only. Hidden Tau evaluation criteria are available only to Tau's offline evaluator after the trajectory is complete.

Source and Model Locks

  • Tau: sierra-research/tau2-bench tag v1.0.0
  • Commit: 17e07b1da2bbc0cadfddeea36412686e0604127b
  • Actor and official user simulator: gpt-5-mini-2025-08-07
  • Proxy: gpt-5-nano-2025-08-07

The repository includes a SHA-256-locked import-only patch that prevents Tau text-mode imports from requiring optional voice and knowledge dependencies. verdict-bench doctor verifies both the commit and the exact working diff.

No-Spend Safety

doctor, freeze, report, and export-tau make no model calls.

smoke and pilot refuse to start unless both conditions are true:

export VERDICT_BENCH_FUNDED=1
export OPENAI_API_KEY=...

They also require a positive --credit-budget-usd. Smoke is capped at $10 in credits. Pilot is capped at the lower of $100 or 20% of the declared awarded credits. This project authorizes $0 of cash-funded API spend before credits are confirmed.

Commands

verdict-bench doctor --fetch-source
verdict-bench freeze --fetch-source
verdict-bench smoke --credit-budget-usd 10
verdict-bench pilot --stage tune --awarded-credits-usd 1000 --credit-budget-usd 100
verdict-bench pilot --stage confirm --tuning-run PATH --awarded-credits-usd 1000 --credit-budget-usd 100
verdict-bench report RUN_DIR --provider-billed-usd AMOUNT
verdict-bench export-tau RUN_DIR

Fixed Pilot Split

Seed 300 selects 8 airline and 16 retail base tasks. airline:49 and retail:113 are development-only and never enter pilot or adjudication. All remaining airline and retail base tasks are reserved for adjudication.

Reporting Boundary

Pilot reports include success with 95% intervals, token classes, gross price-sheet cost, exploration cost, paired bootstrap savings, proxy audit calibration, T8 statuses, refusals, retirements, task failures, completion, and dashboard meter reconciliation.

Pilot output can only recommend whether adjudication may be frozen. It can never issue K-R1 PASS, FAIL, or VOID. Negative and refused outcomes remain publishable under the no-suite-shopping rule.

Synthetic Trace Explorer

The release includes a static explorer generated from the public [SIM] artifact. It exposes candidate generation, proxy outcomes, call-level tokens and spend, certificates, and refusals. Every synthetic status is visibly marked [SIM].

Verification

python -m unittest discover -s tests -v
python -m benchmarks.t5r.fixtures

CI checks source locks, golden parity vectors, split determinism, billing, schemas, budget breaking, hidden-criteria isolation, report reproducibility, the static explorer, secret scanning, and an offline Tau mock trajectory through the official runner.

Limitations

  • No measured Tau result is included in v0.1.0.
  • The T8 modulus L=1 is declared, not empirically proven.
  • Seeded Bernoulli audit calibration is not calibration to hidden Tau criteria.
  • A Tau leaderboard submission from this code is a custom modified-scaffold submission, not a standard verified entry.
  • One external reproduction is required before the result is described as externally validated.

License

Apache License 2.0. Tau remains subject to its own repository license and terms.

About

Certified branch/tool gating evidence harness for public agent workloads

Topics

Resources

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages