AIM on one page.

What Adaptive Intelligent Measurement evaluates, what it costs, what has passed its gate and what has not yet. If your evaluator sent you this, here is the short version. Every line is on the site with its status.

Watch it measure

The question behind an eval bill.

If a score moves and no one can say why, the release waits or ships on a guess. That cost is yours to name. This page shows what AIM can settle today.

Schematic, no data. A score line steps down, with a question mark: the score moved, who moved it? Three suspects. The questions: locked, the same questions every run, ruled out. The judge: measured, and its harshness removed, ruled out. The model: what is left, the change points here.

What it is.

A score can change because the model changed, the judge changed, or the questions changed. AIM measures the judge and locks the questions. What is left points at the model.

Simulation

Set to an error bar of 0.20. Delivered at 0.21 once judge harshness is corrected at each level.

Simulation: 15,000 decisions on judge data from our own runs, seed 20260908. With harshness left in, the same run came in at 1.16. Error bars in logits.

Three error bars drawn to scale against the error bar set in a simulation, 0.20. Judge left alone: delivered at 1.16, about six times the error on the label. One average subtracted: 0.79, better in the middle and wrong at the ends. Corrected at each level: 0.21 in that simulation, back on the label. Simulation on judge telemetry from our own runs, 15,000 decisions, one standard error in logits.

Simulation on judge telemetry from our own runs. The 0.07 above is one judge's gate record of 2026-08-22, unit one judgment; these error bars are a different run, in logits.

What it changes.

If your metrics do not move together, an average can hide the one that fell. Each dimension keeps its own error bar, so the one that fell stays visible.

In build.

Schematic, illustrative values, not measured. One number: the average before and after is nearly the same; it looks steady. Each dimension with its own error bar: A, B and C rise a little; D falls well beyond its error bar.

What it costs.

If the evaluation bill gets argued, the stops are yours to set. The widest error bar you accept. The most questions. The most spend.

You set these stops on every scoped engagement. Automatic routing between judges is in build. Each fee is set on a 20-minute scoping call; see how buying works on the Pricing page.

See pricing

Schematic, no values. Three stops, you set each one: the widest error bar you accept, the most questions, and the most spend, each drawn as a slider with a stop you place. The run stays inside the stops you set.

What has passed, and what has not yet.

The reasoning ruler: Simulation.
Hallucination and persuasion scales: gate not yet passed.
Judge hand-off and stops: in build.
Per-dimension profile: in build.
The same model measured on two dates: not yet shown.
Ruler stability over months, and a validity study: not yet shown.

Schematic. People and AI models, read on the same levels: both point into the same scale, marked 0 to 16. Levels 8 to 12 are shaded: the precision gate passed for the reasoning ruler, error bar 0.07 of a level or better. The gate covers the ruler, not people. Built for both; no performer is placed here.

The precision gate passed for the reasoning ruler. It does not cover people: no person has been placed on it yet.

What we will not claim.

AIM implements at measurement grade the latent-trait/GLMM approach that NIST AI 800-3 highlights as a promising foundation for AI evaluation statistics. AIM is not a NIST standard. It claims no endorsement.

No customer names yet.
We do not publish a cost of getting this wrong. We have not measured yours.
No claim about your own prompts or your own task mix.

The Mona Lisa, undistorted: the calibrated measure — True imageReferenceThe calibrated measure

The Mona Lisa, stretched: stands for uniform leniency; correctable — StretchedCorrectableUniform leniency

The Mona Lisa, warped as in a funhouse mirror: stands for hallucination; exclude, cannot be corrected — HallucinationExclude: cannot be correctedFunhouse mirror

Every judge bends the picture. AIM undoes the bends it can measure and stops using a judge where it cannot. Illustration, an analogy. Painting: Leonardo da Vinci, Mona Lisa, public-domain scan (C2RMF) via Wikimedia Commons.

What it takes to run.

Where the data goes. Hosted plans run on our servers. The self-hosted harness keeps your data on your side.
What your engineer has to run. To be confirmed.
Support contact and response time. To be confirmed.

What happens next.

Watch a recorded session together. Then talk to us about a proof engagement on your hardest case.

Email matt@xlnc.co to set up the call.