Spectralgraph
Banks & credit unions

Independent AI model testing, built for the model-risk file.

When a language-model assistant shows up anywhere in a bank's operation, someone eventually asks for the testing evidence. Today there is almost nowhere to find it. Spectralgraph is an independent AI-testing laboratory that measures fabrication risk in the model you host and hands you a signed, dated report your governance file can point to.

Request the sample report See the public index
FIXED PRICE · FIVE BUSINESS DAYS · NO MEETINGS, EVERYTHING IN WRITING · NO CUSTOMER DATA
The regulatory moment

SR 26-2 left generative AI out of scope. Your examiner still has questions.

Regulators replaced the old model-risk guidance in spring 2026 and formally scoped generative AI out. In practice, examiners are already applying the same principles to language-model tools by analogy: inventory the models, understand their limitations, show your validation evidence. Nothing requires an independent measurement, and we will never claim otherwise. But institutions that answer locally tend to want their own testing evidence on file rather than someone else's word for it.

That evidence gap is what the measurement fills. It's not a checklist or a vendor questionnaire. What it does is measure the model itself, reading the model's own internal signals rather than grading its response output. Using your real question types on a model you host, it maps where the model fabricates, how often, and on which topics, down to specific flagged answers your team can verify with their own eyes. The deliverable is a signed, dated screening measurement built for your governance file. In testing so far it catches fabricated answers at about 0.85 AUC on models it has never seen before; that figure is provisional until the work is published.

What it needs from you

No customer data. Nothing installed. Five business days.

The test never touches customer data. It needs a sample of your real question types, the kind your staff or customers actually put to the tool, and a model endpoint you designate. That includes self-hosted open-weight models and private Azure OpenAI or AWS Bedrock deployments your institution controls. Vendor-embedded AI that the vendor attests to is out of scope, and we will tell you so honestly.

The first engagement establishes your baseline. Most of the value compounds from there: quarterly re-measurement builds a trend record your governance file can point to as models change.

Terms

Fixed fees, written engagement, nothing success-based.

Fees are fixed and never success-based: the LLM Validation Report is $9,500, Continuous Assurance is $6,500 per quarter for one model, and the pre-deployment Model Selection Study runs $4,500–$7,500 by candidate count. Full details and the founding client program are on the home page. Every engagement runs under a written agreement before any work begins.

Straight answers

Questions bank risk teams ask first

Does SR 26-2 require this measurement?
No, and we will never tell you otherwise. The revised guidance formally leaves generative AI out of scope. What happens in practice is that examiners apply the same model-risk principles to language-model tools by analogy, and institutions find it easier to answer those questions with measured evidence on file than without it. This measurement is something a bank chooses, not something a regulator requires.
What does the test need access to?
Almost nothing. It never touches customer data. We need a sample of your real question types and a model endpoint you designate, and everything you share is held confidential under a written agreement before any work begins.
We only use vendor-embedded AI. Is this for us?
Probably not yet, and we will say so. The measurement applies to models your institution hosts or controls, including private Azure OpenAI or AWS Bedrock deployments and fine-tunes. If your only AI is embedded in vendor software and the vendor attests to it, there is nothing of yours to measure, and we would rather tell you that than sell you something.
What about confidential supervisory information?
We never ask for or accept examination materials or supervisory correspondence. The engagement is designed so it does not come near CSI.

Judge the deliverable before spending a dime.

Reply and we will send the sample validation report, the same document your governance file would hold, so you can see exactly what you would be getting.

Email the lab
REPLIES SAME DAY · EVERYTHING ANSWERED IN WRITING, ON THE RECORD