An independent measurement laboratory, built to be checked.

Single-pass measurement of the model's own signals.

The measurement reads a model's token-probability geometry in a single forward pass, on the exact configuration you deploy: same model, same adapters or fine-tunes, same quantization. A documented behavioral protocol runs alongside it, scored against ground truth on your real question domain. Performance on fabricated-entity detection is about 0.852 AUC, length-controlled, leave-one-model-out across 18 models, published in the method paper (DOI: 10.5281/zenodo.21365654). The method came out of ultrasonic nondestructive testing, which is how you test the integrity of a part without opening it.

The failures are printed with the findings.

The lab publishes its methodology, its calibration approach, and its negative results. The validation record includes a pre-registered, hash-frozen experiment whose thresholds were not met, reported in full. The public 18-model reliability index demonstrates the method; client reports are confidential and never published. Every report states what was measured, how, with what error rates and limits, and quotes the specific flagged outputs so your reviewer can check each finding independently.

What the published paper establishes.

The method paper (Logits Don't Lie, DOI: 10.5281/zenodo.21365654) validates a three-stage inspection system across 18 models from 9 providers: recognition measured before a single word is generated, generation monitoring on three temporal signals, and aggregate scoring. The stages fail in different places by design, and combined coverage held on all 18 models tested. We know of no other evaluation framework that measures the pre-generation phase.

Two properties matter if you are relying on the number. The measurement resists coaching: context injection changes what a model says, not the weight-level geometry the score reads, and the injection experiments and their boundary conditions are printed in the paper. And the external signal lines up with what internal interpretability sees. Anthropic's published circuit-tracing work describes a "known entity" feature behind hallucinations; the reading phase measures the corresponding signal from the outside, and the paper documents the two as complementary. The full published record is three papers: Logits Don't Lie, the method paper (DOI: 10.5281/zenodo.21365654); The Doodle Residual, where the measurement began (DOI: 10.5281/zenodo.19283569); and The Transparent Hypercube, the geometry underneath it (DOI: 10.5281/zenodo.19326718).

The fee is published, and it never depends on what the measurement finds.

The lab holds no interest in any client's product, sales, coverage, or certification outcome, and declines remediation consulting for systems it measures. Reports belong to the client, are held confidential under written agreement, and are never published or attributed without written permission. Where an engagement includes reviewer support, technical questions from your designated reviewers get a written answer, on the record, at contact@spectralgraph.ai.

What reviewers ask about the lab