The Value Engine Benchmark: An Evidence-Graded, Methodology-Controlled Environment for Multi-Touch Enterprise-Sales Negotiation

This record contains the pre-results preprint (26 pages) and a snapshot archive of the public repository. The benchmark comprises a fully synthetic multi-touch enterprise-sales negotiation environment with an evidence-graded, methodology-controlled deterministic harness; the frozen evaluation grid covers 13 models and 3,510 graded episodes, with machine-generated paper statistics, dataset datasheet, SHA-256 checksums, and a CI drift gate. The pre-registered judge–human agreement study (Cohen's κ and Krippendorff's α, author-independent raters, frozen sample and thresholds) is not yet run; no human-agreement claim is made in this version. Licensing: the manuscript text is CC BY 4.0; source code in the archive is Apache-2.0; dataset files in the archive are CC BY-NC 4.0. Current code and leaderboards: https://github.com/rudycelekli/the-value-engine-benchmark and https://thevalueengine.ai/benchmark