Spectralgraph
The sample report

Judge the deliverable, not a brochure about it.

This is a complete LLM Validation Report produced by the production instrument for a fictional composite client, Meridian Mutual Insurance. Same pipeline, same report generator, same honesty rules as a paid engagement. Read it before you spend a dime or a meeting.

Open the sample report (PDF) Ask a question about it
REAL INSTRUMENT OUTPUT · FICTIONAL CLIENT · CLIENT REPORTS ARE NEVER PUBLISHED
How to read it

Five sections, in the order a reviewer needs them.

Results summary. The measured fabrication behavior of the model, stated plainly, with the baselines it was measured against.

Per-domain reliability. Where the model holds and where it slips, by question domain, so review effort goes where the risk is.

Flagged answers. The specific outputs the instrument flagged, quoted, so your team can verify every finding with their own eyes. Nothing asks to be taken on faith.

Machine-readable findings. A findings file accompanies every engagement so your governance tooling can carry the evidence forward.

Methods and limitations. What was measured, how, with what error rates, and what the measurement cannot see. This section is why the rest of the document is credible.

The instrument behind it

A screening measurement of the model itself.

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 marketing claims, using your real question types on a model you host or control. 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. The engagement runs in five business days, needs no customer data, installs nothing, and every question is answered in writing.

Straight answers

Questions people ask about the sample

Is Meridian Mutual a real client?
No. Meridian Mutual Insurance is a fictional composite built for demonstration. The measurement, the report generator, and every table in the document are the real production pipeline; only the client is invented. Real client reports are confidential and are never published.
Will my report be published like this one?
Never. Client reports are confidential and client-owned. Spectralgraph never publishes a report or names a client without written permission. This sample exists precisely so you can see the deliverable without anyone's file being exposed.
What would be different in my report?
Your question domains, your model configuration, your flagged answers, and your baselines. The structure is the same: results summary, per-domain reliability, the specific flagged outputs for line-by-line verification, the machine-readable findings file, and the methods and limitations section that states plainly what the measurement cannot see.
Why does the sample admit weaknesses?
Because the honesty is the product. A screening measurement that cannot name its own limits is not one a compliance reviewer can rely on. Every report states its scope, its error rates, and what stays outside them.

Want this file, with your model's name on it?

Send your question types and a designated endpoint, and the signed report is in your governance file five business days later. Pricing is published, fixed, and never success-based.

Email the lab
REPLIES SAME DAY · SEE PUBLISHED PRICING