Whitepapers
Research and analysis
We publish our methodology, our limitations, and our regulatory reasoning alongside our results. That is the standard we argue the industry should be held to.
Regulation / Deployment
June 2026
Inference Under the EU AI Act: Why the Rulebook Points Toward Local LLMs
Europe's risk-based AI regime never says where a model must run, but its demand for control, not merely residency, quietly reshapes the answer. Covers the high-risk obligations, the GDPR and CLOUD Act overlay, and why lossless compression keeps the system your conformity assessment describes identical to the system actually running.
Read online →Legal AI / Compliance
July 2026
Explainable AI for Legal Advice: Why Retention Beats Retrieval for Compliance
Nearly every legal AI product is built on retrieval, and retrieval cannot show what the AI considered when the answer is challenged. Our controlled study on the standard legal RAG benchmark found whole-document reading tripled exact-clause accuracy, and improved it tenfold on merger agreements. Includes the regulatory mapping (EU AI Act, SEC, malpractice), and why the retained record must itself be governed as regulated data.
Read online →Agentic AI / Financial Crime
August 2026
Retaining Agent Working State, and Restoring It Later
What it costs to keep an agent's working state and load it back: lossless compression at 1.444x on bf16, and a 5,097-token session restored in 56.2 milliseconds against 417 to re-read the document, measured on Qwen2.5-7B on an L40S. Sets out what the EU AI Act, SR 11-7, DORA and GDPR actually require of a retained state, and states plainly that a retained cache is numeric working state and not a readable chain of reasoning.
Read online →Infrastructure / Architecture
August 2026
Why the Serving Layer: Building Governed Retention Inside vLLM, SGLang, and NIM
The model's working state lives inside the serving engine and nothing above the API can touch it, so that is where governed retention has to be built. This paper explains why we chose the engines' documented extension points over a fork or a hosted API, why performance demands GPU-native handling, and what the throughput engines and offload stacks do not do: lossless compression, signing, and governance. Includes results from our August 2026 full-precision study, where all 500 benchmark answers were unchanged under compression.
Read online →Reproducibility / Benchmarks
August 2026
The Variance Trap: Run-to-Run Reproducibility of LLM Reasoning, Local and Cloud
Run the identical seeded, temperature-zero evaluation twice and how much does the score move? On a local open-weights 31B, scores moved by at most two points and 2 of 200 answers changed. Through a frontier API, 11 of 200 answers changed, half of all reasoning steps differed, one run was served by seven backend configurations, and a score moved seven points in minutes. Identical scores can even hide churn: answer flips that cancel out. Covers why this is nobody's defect, and why it means the reasoning behind an answer must be retained at run time or it never exists.
Read online →Benchmarks / Evidence
August 2026
Same Answers, Compressed or Not: A Cross-Model Study of Lossless Retention
If you compress the model's working state, do the answers change? We ran the same 500-question legal benchmark across two model families, two sizes, and two precision levels, with and without Revyzor. At full precision on a 31B model, all 500 answers were unchanged; on 200 chained multi-step answers, parity held on the workload where divergence would compound. Includes the runs where 1-2 answers differed, and why that is serving-engine noise, not compression.
Read online →Regulation / Sovereignty
August 2026
The EU AI Act Does Not Mandate AI Sovereignty: A Briefing for Financial Institutions
The Act regulates roles, risk tiers and evidence, not geography, and contains no data-localisation requirement. The obligations that actually bear on where a workload runs sit in DORA, GDPR and supervisory expectations, and they are already in force. Sets out the amended timeline after the Digital Omnibus, including the deferral of Annex III high-risk (credit scoring, life and health pricing, HR) to 2 December 2027, the conditions under which a deployer becomes a provider, and a decision framework for separating sovereignty from theatre. Position as at 12 August 2026, against consolidated text CELEX 02024R1689-20260727.
Read online →Full methodology, per-corpus tables, and benchmark evidence packs are available under NDA. Request access.