Research · Methodology

How we measure


The standard behind every number we publish — for models on our benchmarks, and for the expert work our datasets are built from.

Expert-authored, adversarially reviewed

Every benchmark task and assessment scenario is written by practising professionals, then attacked before it is ever used: an independent model attempts to refute each item — impossible arithmetic, criteria that only reward one school of thought, planted flaws that are actually defensible readings. What survives goes to a named human practitioner who approves or rejects it with their name on the record. Nothing examines anyone until a person has stood over it.

Rubrics with signed weights, not vibes

Work is graded against fixed criteria a grader marks met or not-met, each carrying a signed weight. Credits reward reasoning that was actually done; debits fire only on failures that actually happened. Criteria are written so that a well-reasoned opposite conclusion can still earn full credit — we measure judgment, not agreement. Criteria are fixed to the scenario, never adapted per candidate, so adaptive interviews change the experience without changing the measurement.

Humans anchor the grading

Deterministic checks are automated; judgment criteria are graded by people. As model judges take on grading volume, they are validated against a blind human-graded subset — a judge that cannot match human graders on the overlap does not grade alone. Every grader's verdict is stored per criterion, so agreement is measurable rather than asserted.

Uncertainty is part of the number

Published scores carry confidence intervals and sample sizes. A benchmark result based on too few tasks is labelled as such rather than ranked as if it were settled.

Integrity by design, decided by people

Submissions pass through integrity tripwires designed so that automated tooling reveals itself. A tripped wire is evidence, never a verdict: a person reads the work before anything is decided, because legitimate tools — screen readers, translation — can trip the same wires. Assessment interviews are proctored, spoken, and on camera.

Data quality the same way

The datasets we deliver are built under the same discipline: blind double-annotation with inter-annotator agreement thresholds, gold-seeded quality sampling, human adjudication that feeds back into gold, and per-item provenance — who did what, under what clearance, with the conformity fields EU AI Act deployments need.

Sovereign by construction

The entire pipeline — identity, data, inference, voice — runs on EU infrastructure with no third-country vendor in the data path. That is not a hosting preference; for regulated European buyers it is the product.

Questions about the methodology, or interested in authoring for a benchmark in your field? hello@aisceal.com