Send a model endpoint or an agent scaffold, choose the suite and the effort levels, and get back a VulcanBench report on your own results: private by default, in the same journal format, graded by the same deterministic hidden tests that grade everything else. No LLM judge, no rubric, no special-casing: your model runs the identical harness in the same network-isolated sandbox as the frontier.
Your results in the same format as the public reports: model card, findings, per-task table, and the full PDF. Yours to keep, share internally, or publish.
Every trace, per-task result, cost, and replay transcript: the raw run output behind the summary, not just the headline numbers.
Pick the suite, the effort levels, and the number of repeats. Measure a base model, a fine-tune, or a whole agent scaffold, yours or ours.
We agree on the models, suite, effort levels, and repeats. You get a quote for the compute plus engineering time up front, no surprises, the same cost transparency as the public reports.
Your model runs the identical harness in the network-isolated Docker sandbox: the same tasks, the same hidden tests, the same deterministic grading. Nothing about a paid run is graded differently.
You get the private report and the complete raw data, typically within a week, plus a walkthrough of where your model wins, where it burns cost, and which tasks separate it from the frontier.
Pricing scales with scope: suite size, how many models, how many effort levels, and how many repeats. A focused single-model engagement starts in the low four figures; a broad multi-model sweep with repeats is more. The compute is real and metered, you see it, the same way the public reports show theirs.
Commissioned runs are private and labeled. A paid evaluation never appears on the public leaderboard unless you choose to publish it and it was run under the identical public methodology. Money buys the measurement and the report, never placement, and never a favorable grade. That line is the whole reason a VulcanBench score is worth paying for.
Whether you're ready to scope an engagement or just want to know what a run of your model would cost, start a conversation: a couple of sentences about the model and what you're trying to learn is enough to come back with a plan and a number.
