freightutils-eval: frontier language models on 340 freight reference questions — alone, with the FreightUtils tools, and with web search

A reproducible measurement of how often frontier language models answer freight reference questions correctly on their own, with the FreightUtils tools (MCP and REST), and with a provider's web search in place of the tools. 340 questions in seven families — ADR Table A values, what an ADR Table A code means (a packing instruction limit, a special provision threshold, a tank code), the ADR 2027 amendments, ADR 1.1.3.6 small-load totals, limited and excepted quantities, the HS 2022 hierarchy, freight arithmetic (chargeable weight, CBM, loading metres), and identifiers (UN/LOCODE, AWB prefixes, ULDs) — each with one expected value, scored by a deterministic string rule: no model grades another model. The archive holds the frozen question set (v0.5, sha256 in questions/v0.5/manifest.json; earlier versions kept), the harness, the scoring rule, and for every run the per-question answers, the summary figures and the spend. Licensing: the question set under questions/ is CC BY 4.0; the harness and scripts are MIT (see LICENSE).