(Urs) — Unified Resource Sphere

Your knowledge feeds models that will serve your competitors.

(Urs) preserves it, reveals its singularity, and amplifies what only you know.

No knowledge amplified without knowledge verified. It all starts with proof.

(Urs) — the cognitive trust layer

AI now writes part of what you sign — and no one checks what it claims.

One deliverable. 86 claims.
How many are true?

Evidence audits any AI-assisted deliverable — whether you sign it or receive it. Every claim typed, sourced and scored 0–100 against 600M+ sources. Deterministic verdicts, replayable years later. Sovereign, hosted in France.

Audit a deliverable Get a sample report

In production · built for Articles 12, 14 and 19 of the EU AI Act

evidence · forensic verificationsession 51e88df6

Claim extracted

“Open-plan offices increase team productivity”

sources 29 peer-reviewed 20 passes 3× self-consistency

Score CRVV 74 / 100

CONTRADICTED

Claim SHA-256 3fa1c88e…91b02e7 · deterministic replay < 50 ms · immutable audit chain · EU AI Act Art. 12

claims extracted

hallucinations detected

1 401

unique sources mobilized

8 min

end to end

Real run — AI-generated banking brief, drafted by a frontier LLM, audited before internal use · June 2026 · full report available on request

Both sides of the table

Before you sign it.
Before you pay for it.

A large language model predicts the most probable word, not the true one. It fabricates — with confidence. The same verification protects the one who produces and the one who commissions.

You produce it

Before you sign it

Consulting firm, research department, advisory team: AI makes your deliverables ship faster — and your signature vouches for every claim. Evidence intercepts fabrications before they ship, and embeds the proof of verification in the document. Your reputation is no longer one hallucination away.

You commission it

Before you pay for it

Procurement, public buyer: in 2025, a 237-page report delivered by a Big Four firm to Australia's federal employment department, invoiced at A$440,000, cited academic work that does not exist and attributed to a federal judge a sentence he never wrote. A few minutes of verification would have taken down the first citation. Demand the verdict before acceptance.

The same forensic report serves as proof for one side and acceptance test for the other — deterministic verdicts, cryptographic fingerprint, replayable years later.

One layer · three doors in

Enter through proof.
Stay for memory.

Evidence — the entry wedge

Forensic audit of your deliverables

Any AI-assisted deliverable, audited before it ships or before acceptance. Every claim typed at extraction — opinions and predictions are shown, never scored: Evidence doesn't manufacture certainty on what can't be verified. Three-level verdict, deterministic, replayable.

Pilot footprint3 months · cancel anytime

Start a pilot

Reveal

The cognitive MRI, in 5 days

In Deloitte's Global Human Capital Trends study, 75% of organisations rate preserving their knowledge as important or very important; 9% consider themselves fully ready to do so. Reveal measures that gap in your organisation, live, on your corpus: hidden experts, knowledge no one else holds, blind spots. No slides — a map your executive team queries directly.

Committed diagnostic5 days · fees deducted from Cortex kick-off

Request a Reveal

Cortex — The Serendipity Infrastructure™

The sovereign platform

Your knowledge preserved on your servers, amplified with 600M+ qualified world sources, exposed to your teams as an operational memory — every interaction traced, audited, reproducible. And above all, serendipity: Cortex answers your question, then reveals what you didn’t ask — the connections between your knowledge and the world’s that no team sees alone.

Annual platformon-premise or European sovereign cloud · air-gap capable

Discuss Cortex

The platform, in production

Not a mockup.
A runtime.

cortex · cognitive signature — runtimedemo tenant · redacted

An organization's cognitive signature — living runtime, names redacted

Multi-source verification reportaudit · e77e3a58

Critical risk

This document presents significant risks.

213 / 410

verified

to arbitrate

hallucinations

contradictions

59/100credibility

82/100relevance

66/100verifiability

68/100veracity

410 claims · ≈ 10 s per claim, end to end · 6,035 unique sources incl. 2,573 peer-reviewed · 3 self-consistency passes · SHA-256 c9e0994c…07d72a0 · deterministic replay — EU AI Act Art. 12

Evidence report — real run, June 15, 2026, on a 410-claim banking expertise dossier

Human arbitration zones — NUANCED doctrine83 documented

#366 ✗ Hallucination detected

“Potential across the top 200 banks: +5% revenue, −8% co…”

None of the 6,035 sources mobilized attests it — remove or rewrite before publication

#66 ◐ Conflicting sources

“European banks hold on average 3.1 points of own fun…”

Qualified Tier-1 dissenting voice — arbitrate per business context, sources on the table

#111 ◐ Conflicting sources

“Online banking penetration reaches ≈ 72% in France.”

Credible Tier-1 contradiction — decide, or surface the nuance in the document

Expert arbitration zone — excerpts from the same run. Uncertainty is never hidden: it is handed to the human.

Sovereignty · the page your procurement will read

Your reasoning never leaves your walls

Generic AI sends your intellectual property to foreign clouds — tomorrow's training data, and a service that can be cut off overnight. (Urs) takes the opposite stance: execution on your servers or on European sovereign cloud, air-gap capable.

EU AI Act · Art. 12

Record-keeping

Immutable audit chain, six fields per event, replayed identically in under 50 ms — six months or six years later, whatever the models deployed since.

EU AI Act · Art. 14

Human oversight

When the evidence is genuinely mixed, (Urs) doesn't fake certainty: it hands the arbitration to your expert, sources on the table. A canonical pipeline step, not an afterthought.

EU AI Act · Art. 19

Retention

Decisions and their context are retained in the code, not promised in a contract. Compliant by construction — retrofitting takes years.

Hosted in France (OVHcloud) or on-premise · SHA-256 integrity per document · model fingerprint persisted · regulation is not our cost line — it is the moat.

Built, not promised

Built. In production.
Third-party attested.

386,000+

lines of proprietary code · 5,600+ methods · 250+ endpoints — measured 8 August 2026

candidate patent families, selected from 26 invention disclosures reviewed — independent Fidal Innovation assessment, filings in progress

24 months

of R&D collaboration with CentraleSupélec Paris-Saclay

“No cluster appears today to represent artificial intelligence from a cognitive standpoint.”
Independent IP assessment — Fidal Innovation, 2026

This cognitive infrastructure with no recorded equivalent has a name: The Serendipity Infrastructure™. It is what the assessment attests — and what eight patent families protect.

The team

Vision and execution,
in the same room

Fabien Lomet

CEO & Founder

Thirty years across consulting, revenue leadership and technology — PwC, IBM, ServiceNow, Kyriba. Designed and built the (Urs) platform over thirty-six months. Founding thesis published (HAL, SSRN).

Franck Douaud

CPO & Co-founder

Cognitive systems and user experience pioneer. Three decades making complex systems simple for the enterprise.

Scientific advisory board

Frédéric Pascal

VP AI, Université Paris-Saclay

Director of the DATAIA Institute, France 2030 AI-Cluster — 1,000+ researchers. Professor at CentraleSupélec, appointed expert at the CSPLA.

Alexandra Bensamoun

Led France's AI Act mission

She doesn't just advise on regulation — she shapes how it applies. France's official AI Act mission (CSPLA, Dec. 2024), French AI Commission member.

Claire Gardent

French Academy of Sciences

Elected to the French Academy of Sciences (June 2026). CNRS Silver Medal, ACL Fellow, research director at Loria — leads the AI ENACT chair.

Sébastien Laye

Entrepreneur & Economist

Advises European tech leaders on US market entry and scaling. Lecturer at HEC Paris — “AI as your co-worker”.

Research & notes

What we publish

Essay · serendipity

Industrialising serendipity

Half our pharmacopoeia was born of accidents seen by the right person. Yet AI in 2026 optimises prediction and is becoming “less tolerant of open-ended experimentation”.

Founding essay · from the thesis · HAL · SSRN

From extraction to preservation

Resources, labor, attention… and now cognition. The genealogy of the paradigm shift — cognitive phenotypes, sovereignty, and why amplification will replace disruption. 11 min.

Essay · agent governance

Identity is not mandate

Your agents no longer answer: they act. July 2026 showed what is missing — not safer models, but a written mandate, factual grounding and replayable proof.

Essay · reliability

The prompt that makes AI honest does not exist

Seven rules to paste into the settings, and AI would stop lying. A calibration failure cannot be fixed by an instruction — reliability is built outside the system it verifies.

Method · NUANCED doctrine · upcoming

Why three verdicts, not two

What a binary verdict misses in a real deliverable, and where the human must decide.

Note · Digital Omnibus · upcoming

AI Act: high-risk (Annex III) deferred to December 2, 2027

Prohibited practices (Feb. 2025) and GPAI (Aug. 2025) already in force, transparency in Aug. 2026; Annex I by Aug. 2, 2028. The Omnibus deferral buys time — and the regulator says prepare now.

All publications, 2024-2026 →

Founding thesis — (Urs)'s theoretical stance: cognition is not the next extraction frontier. Full text, open access: HAL → · SSRN →

Technical annex

For your IT, compliance and security teams

The mechanisms in detail, separate from the pitch. Each block opens independently.

Positioning — on top of the model, never in its place

(Urs) sits on top of any engine — it doesn't replace it

(Urs) — the cognitive trust layer

Sovereign by design · France (OVHcloud) or on-premise · air-gap capable

Sovereignty

Reasoning never leaves your walls

Verifiability

CRVV + replayable audit chain — EU AI Act

Memory

Knowledge persists beyond the model

Serendipity

Cognitive chemistry connects the unseen

Interchangeable engines — your choice, swappable anytime

Mistral

Claude

Gemini

ChatGPT Ent.

Powerful on its own — but four gaps no model closes alone

hallucinates — answers without provable sources

forgets — no organizational memory

leaks — your reasoning trains someone else's model

can't prove — no audit, no replay

(Urs) makes any LLM sovereign, verifiable and yours

Trust scoring CRVV · 0–100

Four dimensions — credibility, relevance, verifiability, veracity — measure the documentary quality of sources, deterministically: the same input yields the same score, run after run. This is what allows an audited decision, where an LLM's confidence score varies between runs. Industry presets (pharma, finance, consulting, tech, regulatory) that each client refines into its own trust policy.

Reasoning audit RCQ · reasoning chain quality

CRVV audits the sources mobilized. RCQ audits the reasoning that produces the verdict — step elaboration, cross-pass self-verification, hypothesis exploration, granularity adapted to local complexity. A correct verdict produced by flawed reasoning remains a silent risk: RCQ makes it measurable and auditable. What no LLM-as-a-service tool audits. Calibrated by expert human annotation, scientific publication in progress.

Claim typing extraction · verdicts

Every claim is typed at extraction. Non-verifiable ones — opinions, predictions, rhetorical statements — are shown but never scored. Verifiable claims receive a three-level verdict: VERIFIED, TO ARBITRATE, CONTRADICTED. In grey zones (5–15% of a typical deliverable's claims), the decision goes to your expert with the sources: stance, weight, credibility, full text. Academic and business coverage.

Replayable audit chain forensics · replay

Six traceability fields per event, SHA-256 fingerprint per document and per claim, model fingerprint persisted. Replay reconstructs any decision identically in under 50 milliseconds — six months or six years later, regardless of models deployed since. The final report embeds the source document and the full forensic audit as attachments — a third party recomputes the fingerprint and verifies the verdict without any platform access.

Qualified world corpus 600 M+ sources

Over 600 million peer-reviewed academic sources (OpenAlex, CORE, PubMed, HAL, Semantic Scholar) and validated open-science sources, complemented by a third tier: the qualified web, curated in real time. Every source is scored before it serves — never a raw web crawl. These sources enrich your corpus without ever compromising it.

Deployment & models sovereignty

Three modes: on-premise on your servers, European sovereign cloud (OVHcloud), or fully isolated — complete air-gap. (Urs) sits on top of any model without replacing it: Mistral, Claude, Gemini, ChatGPT Ent., other — swappable anytime. Your reasoning never trains a public model. Platform access by invitation.

NOTE — Certain internal mechanisms (knowledge structuring into persistent units, semantic bonding engine, pipeline orchestration) are deliberately not detailed on this page: eight patent families are being filed with Fidal Innovation. Details available under NDA.

Start seeing

Your next deliverable, audited.
Or your corpus, revealed.

An Evidence pilot on a real deliverable — or the 5-day Reveal on your infrastructure. Either way: you see what you know, for the first time.

Audit a deliverable Get a sample report Talk to Fabien

or write to us directly: contact@urs.ai