Inspect the files. Recompute the findings.

Regrade all recorded answers

Python 3.11+ and Node 20+ are required. Install the small review dependency set, then reproduce the uniform formatting correction and published scores.

git clone https://github.com/levco/crebench.git
cd crebench
python3 -m pip install -r requirements-review.txt
python3 tools/collect-workflow-results.py experiments/2026-09-06-workflow-v1
python3 -m unittest discover -s tests -v
npm run build
npm run check:site

Recalculate actual Excel files

Install LibreOffice and make soffice available on PATH. The public replay checks original workbooks and the published changed-input copies against independent financial expectations. It verifies each edited cell and records your local engine version.

python3 tools/replay-workflow-audits.py experiments/2026-09-06-workflow-v1 \
  --output work/independent-replay

The scored run used LibreOfficeDev 26.8.0.0.alpha0. Engine differences are recorded, not hidden. The replay reads published XLSX files without rewriting their formulas.

Run the models again

The frozen runner, complete prompts, tool schemas, context transforms and model IDs are published in crebench/. A new paid run must use a new experiment directory and disclose its own model availability, version and date.

Reproduction limit: The scored general-agent artifact writer used the Codex-bundled @oai/artifact-tool package, which was not available from the public npm registry when checked. Its wrapper source is public, but identical artifact generation currently requires that runtime. The public replay and numerical grader have no dependency on it. Lev’s native backend remains proprietary; its original outputs, operating conditions and sanitized review evidence are public.

Challenge a finding

Include the case, system, exact criterion and supporting page or cell. Author reviews are open to disagreement. Cite the original output and show the proposed correction; every affected system must be regraded together.