Spectralgraph
AI vendors

Your buyers stopped taking your word for it.

Enterprise security questionnaires grew an AI section. Hospital committees ask for validation evidence. Underwriters want a testing file. Spectralgraph is an independent AI-testing laboratory that measures the model behind your product and hands you a signed evidence pack built to be shown to the people deciding whether to buy from you.

Email the lab Read a sample report
$12,500 PUBLISHED · FIVE BUSINESS DAYS · YOUR REPORT, YOUR CONTROL · FEES NEVER CONTINGENT ON RESULTS
The problem

Deals stall in security review because self-reported evals stopped counting.

The AI sections now standard on enterprise questionnaires ask for model provenance, hallucination controls, and evidence that outputs are evaluated. An internal dashboard cannot answer them, because you graded your own homework. Reviews that take weeks with nothing independent to show tend to take days with a signed third-party measurement attached.

The Vendor Evidence Pack

One measurement, reused across every deal for a year.

Evidence report
A signed, dated screening measurement of fabrication behavior in the model behind your product, on your real question domain: where it fabricates, how often, on which topics, with the flagged answers quoted for line-by-line verification, plus methods, error rates, and limitations stated plainly.
Summary letter
A concise signed letter written for your buyers' reviewers, insurers, and auditors. Yours to distribute; never published by the lab; never names you without written permission.
Evidence brief
One page formatted for security-questionnaire responses and trust-center posting, so your sales team never improvises what the letter is.
Written reviewer support
For twelve months, your buyers' security reviewers can put technical questions about the measurement to the lab directly and get answers in writing, on the record. The lab describes the measurement and never advocates for a purchase.

It's not a checklist or a vendor questionnaire. What it does is measure the model itself, reading the model's own internal signals where your stack exposes them and a documented behavioral ground-truth protocol where it does not. In testing so far the instrument catches fabricated answers at about 0.85 AUC on models it has never seen before; that figure is provisional until the work is published.

Also in the line

Two questions your buyers ask that now have measured answers.

Training-Data Ingestion Screen, $8,500 per training event ($4,500 alongside an Evidence Pack): a ratio-based screening measurement of whether your fine-tune absorbed a specified corpus, with positive and negative controls. It turns the "we never train on your data" paragraph into evidence. Its limits travel with it: meaningful-content ingestion, not rote strings; screening language, never a certification.

CHAI Applied Model Card supplement, $2,500: measured results formatted into the model-card fields health-system buyers now request, alongside Joint Commission responsible-AI guidance.

Keep the evidence fresh with the Evidence Refresh, $5,000 per quarter (founding vendors lock $4,000 for two years). Full numbers on the pricing page.

Founding program

Three founding vendors. $6,500. Invoiced on delivery.

The first three vendor engagements run at $6,500, invoiced only on delivery, in exchange for permission to name the engagement in a short case study you approve in writing. The rate is low on purpose: the quarterly refresh relationship is the real business, and the first three vendors make it possible. After that, the list price is $12,500 and it does not move.

Straight answers

The five questions vendors ask first

Will our buyer's security team accept a report from a lab they haven't heard of?
The report is built to be judged on its contents: method, controls, error rates, limitations, and the exact flagged answers, all reproducible. It reads like an instrument log, not a badge. And for twelve months your buyers' reviewers can send technical questions to the lab directly and receive written answers on the record. A one-page lab brief ships with every pack so you never have to explain who Spectralgraph is yourself.
What if the results are bad?
Then you will know before anyone else does. The report is confidential and client-owned: the lab never publishes it, never names a client without written permission, and you choose who sees it. Most vendors in that position fix what the measurement found and re-measure within 60 days for a new, separately dated letter. What never happens is a finding being adjusted, because fees are fixed and never contingent on results.
We run on OpenAI or Anthropic APIs. Can you still measure us?
Yes. The behavioral protocol tier measures your product's fabrication rate on your real question domain against documented ground truth, with the full protocol disclosed so your engineers can re-run it. Where your deployment exposes token probabilities, vLLM-class stacks, Azure OpenAI, Bedrock custom imports, the instrument-grade measurement is added on top.
Our engineers already run evals. Why pay for this?
Keep them; they are good tools. The difference is one word: independent. An internal eval dashboard is your own team grading its own homework, and enterprise reviewers now say so out loud. This is the measurement nobody on your payroll produced, which is exactly what the AI section of a security questionnaire is asking for.
Does the measurement need our customers' data?
Never. Probe sets plus a synthetic or de-identified question domain you supply, against an endpoint you designate. Nothing is installed, and everything you share is held under a written agreement.

Send the two facts, get the plan the same day.

Tell the lab what your product is and which deal, review, or application is asking for evidence. You will get a written plan, the engagement letter, and the sample report by reply.

contact@spectralgraph.ai
REPLIES SAME DAY · EVERYTHING ANSWERED IN WRITING, ON THE RECORD