Public Interest

As AI becomes a layer over search, work, education, policy, medicine, and finance, the public needs records of what these systems surface — and what they do not. Without one, that behavior drifts unseen, and the decisions built on it drift with it.

Why records matter

AI answers are ephemeral. A response given today may differ tomorrow, and most users have no record of what was omitted. Imbas preserves observations under documented conditions so patterns can be inspected over time.

Public-interest measurement

Imbas documents observable behavior so researchers, journalists, institutions, and citizens can inspect surfacing patterns without relying on vendor claims or anecdote.

Why tracking this matters

What gets measured gets better. Tracking what AI surfaces — and omits — over time makes the gaps visible, and visible gaps get closed. It also catches drift: a model that volunteers a mechanism today may quietly stop tomorrow, and with no record, no one notices. As AI becomes the layer people decide on — medicine, finance, policy — undetected drift degrades those decisions invisibly, at scale. A public record keeps decisions anchored to what’s true instead of drifting with the model. And it’s tractable: this behavior is observable from the outside, at low cost. Researchers, journalists, and institutions can all inspect it — but the point is simpler than oversight. Better outputs on net, steadier decisions, for users and labs alike.

How it’s measured

v1 established the phenomenon: 25 cases scored on a 0–3 rubric. Those scores were first-pass and founder-assigned — enough to show the gap is real and recurring, not enough to call it settled. v2 adds a blinded sub-study scored by an independent collaborator, so the rubric can be tested for inter-rater reliability. The instrument is being built to be checked, not trusted.

Not a verdict system

Imbas does not declare intent. It does not decide truth for the reader. It does not render moral judgment. It records behavior.

The public-interest value comes from making the record inspectable. A reader can examine the captured response, the rubric, and the cited evidence and reach their own conclusion. That is the point of public records: not to settle the question, but to make the question inspectable.

What public support can produce

Output
Expanded case archive across topics and models
Output
Methodology development and versioned documentation
Output
Public case reports with full prompt transcripts
Output
Public datasets for independent analysis
Output
Field Notes on emerging patterns
Output
Live signal prototype for in-the-moment gap detection

A measurement system that cannot produce small gaps, null findings, and ambiguous results is not a measurement system. It is a conclusion engine. Imbas is not trying to be a conclusion engine. The v1 dataset contains controls that produced gaps as low as 0.75, including one model that scored a perfect 0 on Case 013 (OxyContin). That variance is the credibility.

Support the record

This work is funded to stay independent — outside the labs it measures. Imbas is seeking grant support and partners to expand the archive and complete the v2 reliability study.

Get in touch