Straight answers, on the record.

Fabrication. It measures the model itself, reading the model's own internal signals rather than grading its response output, and it maps where a model fabricates, how often, and on which topics, using your real question types on a model you host. Flagged answers are individually verifiable by your own team.

A sample of your real question types and a model endpoint you designate. It never touches customer, member, policyholder, or patient data, and nothing is installed in your environment. Everything you share is held confidential under a written agreement before any work begins.

Models your organization hosts or controls whose deployment exposes prompt-side token probabilities: self-hosted open-weight models, vLLM-class private-cloud deployments, and your own fine-tunes. Private Azure OpenAI is the exception: Azure exposes token probabilities on output tokens only, not the prompt side the instrument reads, so the instrument can't measure those deployments. Vendor-hosted SaaS AI that the vendor attests to is out of scope.

No. The deliverable is a screening measurement: a signed, dated report of measured model behavior. We are not an audit firm or a certification body, and we do not issue opinions or attestations.

It catches fabricated answers at 0.852 AUC on models it had never seen before: fabricated-entity discrimination, length-controlled, under leave-one-model-out validation, a result published in the lab's method paper (DOI: 10.5281/zenodo.21365654). A public demonstration of the method, the AI Reliability Index, covers 18 open-weight models across 15 professional and regulated domains.

The lab is written-first and async on purpose: every question answered in writing, on the record. Compliance buyers tend to prefer having the answers in the file over having had a phone call about them.

Fixed fees, never success-based: LLM Validation Report $9,500; Continuous Assurance $6,500 per quarter for one model, $8,000 for a second model or a heavy change cadence; Model Selection Study $4,500 for two candidates, plus $1,500 per additional candidate, to a maximum of four for $7,500. A client's first engagement is invoiced only on delivery, under the written guarantee described on the pricing page.

Five business days from materials-complete: your question types received and your endpoint reachable. Replies to email come the same day.

A signed, dated validation report plus the complete findings file: fabrication frequency by topic, the specific flagged answers, and the measured baseline your governance file keeps. Ask for the sample report by email and judge the deliverable first.

No. The EU deferred its high-risk AI obligations to December 2, 2027 (adopted June 2026), and even then conformity is provider self-assessment; nothing in EU law requires an American insurer, bank, or health system to buy independent testing. One footnote: the EU's code of practice for frontier-model makers now expects qualified independent external evaluators, the clearest sign yet that independent AI testing is becoming the norm. The obligations with living dates for US buyers are the US ones: NAIC model bulletin adoptions, the Joint Commission certification, and ABA guidance.

Connecticut adopted the NAIC model bulletin as Bulletin MC-25 and asks every Connecticut domestic insurer to certify annually, on or before September 1, either that it does not use AI systems or that its use is substantially consistent with the bulletin. The certification is the insurer's own statement; nothing requires independent testing. A measurement adds a dated, independent record to stand behind that signature. The NAIC model bulletin crosswalk on this site maps the report to the exact text Connecticut adopted. The Connecticut page lays out the timeline.

Not by coaching it. Context injection changes what a model says, not the probability geometry the measurement reads: in the published injection experiments, prompt-side coaching improved the text while the geometry barely moved, with architecture-specific boundary conditions documented alongside (DOI: 10.5281/zenodo.21365654). The score reads whether trained pathways exist in the weights, so the reliable way to raise it is to train the knowledge in, at which point the model has genuinely improved.

In the lab, yes. The published method reads three stages: a pre-generation phase that reacts while the model is still processing the prompt, before any answer exists; per-token signals during generation; and the aggregate measurement afterward (DOI: 10.5281/zenodo.21365654). The Beyond the Scorecard page shows demos of the first two, built from real measurement data. The difference from the detection work now appearing in research is structural: our instrument is not a trained classifier. There's no labeled dataset behind it and it never touches the model's internal states; it reads the token probabilities a vLLM-class endpoint exposes, and the published validation was run on models it had never seen. What we sell today is the measurement: the signed report and the quarterly trend record. A live version that acts on a detection the way your policy chooses, block it or flag it and let it pass, is in development on the same patent-pending instrument. If in-stream handling is what your deployment needs, write us and say so; that interest shapes the build order.

Both. The reading phase measures whether the model recognizes a topic while it is still processing the prompt, before any answer exists, and the writing phase plus three temporal signals monitor the generated response. The pre-generation signal reaches comparable effect sizes from roughly ten times fewer tokens, and it reads that phase from token probabilities rather than the model's hidden states. Per-model results for all three stages are published (DOI: 10.5281/zenodo.21365654).

Not by default. In the published 18-model index an 8B model outscores a 70B from the same family, because small models fabricate visibly while large models fabricate with the same fluent geometry as genuine retrieval, which makes visible uncertainty a safety feature. Aggregate scores also hide failure shape: two models a dozen points apart can differ in kind, one growing more cautious on harder questions while the other collapses at the frontier tier. The tier-by-tier curves are in the method paper (DOI: 10.5281/zenodo.21365654).

You won't need one. No PHI ever reaches us, so there is nothing for a BAA to cover. Spectralgraph is a no-PHI vendor by design. We provide a written no-PHI design statement instead.

We never ask for it, accept it, or handle it. The measurement needs your question types and an endpoint you designate.

Four commitments, in writing. Every report is signed by name and dated. Any factual or methodological error your reviewers find is corrected and the report re-issued at no cost. Errors and omissions coverage is bound before any client work begins. And every technical question your examiners or reviewers raise gets a written answer, on the record. Fees are fixed and never depend on what the measurement finds.

On your first engagement the invoice arrives only with the delivered report, and it carries a written guarantee: if the report gives your governance file nothing usable, say in writing what is missing within fourteen days and the invoice is cancelled. The guarantee covers the deliverable, such as agreed scope left uncovered; what the measurement finds is never a reason to cancel.

It doesn't decide truth, and it doesn't replace your people. It measures the model's fabrication signals and flags exactly what deserves human eyes. Every report says so on page one.