Technical Whitepaper · May 2026 · G-1 v2.0

Geodesia G-1.
The Enterprise AI Trust Layer.

The runtime safety and compliance layer that sits between any enterprise application and any open-source LLM. Architecture, platform capabilities, and three production use cases in banking, insurance, and multi-agent AI pipelines — written for compliance officers, Chief Risk Officers, and technology leadership preparing for EU AI Act Annex III enforcement from August 2026.

Download Whitepaper (PDF) G-1 Product Page →
EditionMay 2026 · v2.0
FormatPDF · ~520 KB
AudienceCCO · CRO · CISO · Head of AI
StatusConfidential · authorised recipients

A runtime safety layer.
Zero weight modification.

Geodesia G-1 is a runtime safety and compliance layer that sits between an enterprise application and any open-source Large Language Model. It does not replace the LLM. It scores every inference, blocks prompts that present unacceptable risk, flags hallucinated or unsafe answers, attributes causal responsibility to specific tokens, and seals every interaction in a tamper-evident audit log.

G-1 is built around the obligation pattern of the EU AI Act (Regulation EU 2024/1689) and emits evidence reusable across GDPR, the EU Charter of Fundamental Rights, ISO/IEC 42001:2023, NIST AI RMF 1.0, US state AI laws (Colorado SB21-169, NYC LL144), and SOC 2. Each compliance report explicitly maps G-1 capabilities to specific articles of these frameworks.

Read the full document →
01

Hallucination

Base models detect their own fabrications with near-random precision. In multi-agent pipelines a single false figure propagates across every downstream output unchecked. In credit scoring, clinical decision support, or legal drafting, this is not a quality problem — it is a liability event.

02

Regulatory deadline

The EU AI Act begins high-risk enforcement in August 2026, with fines up to EUR 35 million or 7% of global turnover. Seven additional AI laws are already active globally. Every enterprise in a regulated sector has a court date.

03

Sovereignty constraint

Cloud AI APIs require data egress. Banks, hospitals, defence ministries, and public-sector organisations cannot send sensitive records to third-party cloud infrastructure. They are architecturally blocked from the models that would otherwise provide adequate safety controls.

Six chapters.
Architecture to enterprise deployment.

From the problem statement to the runtime architecture, from the compliance platform to three live enterprise use cases. Everything a CCO, CRO, or Head of AI needs to evaluate Geodesia G-1 against an upcoming high-risk AI system deployment.

Chapter 1

Introduction

The Problem · Who This Document Is For · Scope of This Document. The three simultaneous structural failures faced by enterprises deploying open-source LLMs in regulated sectors.

Chapter 2

How Geodesia G-1 Works

Six sequential checkpoints — Request Ingress, Safety Gate, Constitutional Router, Model Inference, NSP Coherence Engine, Compliance Runtime. Plus Data Sovereignty, Constitutional Intelligence, the Compliance Platform.

Chapter 3

Platform Screenshots

Unguarded models — hallucination without detection. G-1 detection and block. Causal explainability — why did the model hallucinate? Side-by-side comparisons from the live dashboard.

Chapter 4

Enterprise Use Cases

Banking — Compliant AI Credit Intelligence. Insurance — AI Claims Pre-Assessment under regulator order. Multi-Agent AI — Pipeline Forensics and credit assignment.

Chapter 5

References

Peer-reviewed academic publications underpinning the detection methods (MuPAX, EVIDENCE, NSP). Legal and regulatory sources mapped by the compliance platform.

Chapter 6

About Geodesia

University of Bari spinoff. Founding team across Bari, Oxford, Stanford, and Uber AI. 10,000+ combined citations, 5 successful exits, 1 IPO, EU patent pending.

Benchmark Highlights

The numbers
behind the trust layer.

The whitepaper documents the full benchmark methodology, dataset, and per-suite breakdown. Below, the headline results: G-1 mounted on Gemma 4 E2B — a model two orders of magnitude smaller than the closed frontier — reaches frontier-class scores on truthfulness and safety, while adding under 35 ms of end-to-end latency.

Download PDF Full benchmark breakdown
0.82
Safety AUROC
+79% vs base
0.96
Hallucination AUROC
+316% vs base
<35ms
Total overhead
end-to-end
13
Frameworks
natively mapped
Made in Europe with

Get the full whitepaper.
Then talk to the lab.

Download the PDF and share it with your compliance, risk and AI leadership. When you're ready for a deep technical walkthrough, sandbox access, or a reference-architecture review, the lab is one email away.

PDF · ~520 KB · 10 pages No form required Confidential — for authorised recipients