Spectralgraph

Spectralgraph is an independent testing laboratory that provides fabrication-risk measurement for language models used by organizations across several regulated sectors, including law firms and legal departments that self-host models for confidentiality reasons.

Key features and functions include:

Fabrication-risk measurement
The core offering analyzes a language model's token-probability patterns during processing and generation to distinguish between fact retrieval and fabrication, without requiring ground-truth answer keys or repeated generations. The method is described by the company as adapted from ultrasonic non-destructive testing techniques and is the subject of a pending U.S. patent application.

Confidentiality-safe engagement design
The measurement process is designed to avoid access to client files, privileged material, or other customer data. Organizations provide representative question types and designate a model endpoint they control; the company states the deployment can include private hosted model environments.

Application to legal practice
For law firms and legal departments specifically, the company positions its measurement as support for the verification duties described in ABA Formal Opinion 512, following court sanctions over AI-fabricated case citations. The engagement is intended to help firms direct verification effort toward the topics and question types where a self-hosted model is most likely to fabricate.

LLM Validation Report
A one-time engagement producing a signed, dated report and complete findings file identifying domains and specific flagged answers where a model's fabrication risk is elevated, intended for use in an organization's governance or risk file.

Continuous Assurance
A recurring quarterly engagement that re-measures a model over time, producing a trend record and including change-gate reviews when an organization swaps or fine-tunes a model.

Model Selection Study
A pre-deployment engagement that comparatively measures fabrication risk across candidate models before an organization commits to one.

AI Reliability Index
A publicly available scorecard benchmarking a set of open-weight language models across multiple regulated domains under standardized testing conditions, published by the company as a demonstration of its methodology.