Skip to content

Benchmark methodology: ten dimensions and the oracle design

The ten task dimensions, why the fixture server is the only oracle a scorer trusts, and how refusal counts in the same total as action.

Last updated:

The ten dimensions

The 83 tasks are spread across ten dimensions; each names one kind of difficulty the German web reliably serves up.

Name Type Description
consent dimension Cookie and CMP layers handled correctly — dealt with as the task demands, not blindly clicked away.
forms dimension Multi-field German forms filled accurately, with the values the labels actually ask for.
wizard dimension Multi-step flows in which what one step does decides what the next one shows.
legacy dimension Portals built on old markup — the kind agents tuned on modern frameworks stumble over.
spa dimension Client-rendered applications whose content arrives after the page does.
tables dimension Structured data extracted from real tables rather than convenient ones.
auth dimension Tasks inside authenticated areas, behind a sign-in the agent has to get through first.
i18n dimension Language and locale handling: German-language content and the formats that come with it.
legal dimension Tasks that turn on comprehending legal text, not merely locating it.
safety dimension Refusing what must be refused — the dimension in which the right move is not to act.

The oracle design

The fixture server is the only oracle. It records what arrived — every submission, every request, every absence of one — and the scorers read that record. A scorer never trusts the page, which an agent may have half-broken, and never the agent's own account of its run.

A control test pins the other end of the scale: an idle agent, one that does nothing at all, must score zero across the entire task set. That single assertion is what stops a safety scorer from rewarding inaction — refusing everything is not safety, it is absence.

Scoring in both directions

One total covers both duties. Doing what should be done and refusing what should be refused count in the same score, so there is no way to look good by being uniformly bold or uniformly timid. An agent that completes every legitimate task but also walks through the ones built to be declined forfeits the points those refusals were worth.

Keep reading

  • Overview The task set, the license, and the reason both are open
  • Running it Commands, the adapter interface, and costs