Xplore Intelligence, with sources.

Company

Name

Xplore Intelligence Ltd — AI research lab

Where

Edinburgh, Scotland · registered in Scotland, company no. SC852183

What

Infrastructure for governing and assuring AI agents: we record what an agent actually did, check what it claimed against that record, and certify the result — for any agent, on any runtime, without modifying it.

Founders

Alexander Shkrebelo, co-founder and CEO · Egor Izgarshev, co-founder and CTO

One line

“We made agent behaviour computable.”

What we make

Two products on one substrate, adopted separately. Most enterprises start with the second.

Govern

Node Resolution and digital twins — a typed graph as the operating system for the agents an organisation builds. The agent proposes; the control loop verifies, commits and journals. A new domain is a configuration, not a system.Validated at TRL 4

Assure

Dome → TrustGraph → Data Integrity passport — capture what any agent saw at a boundary the enterprise owns, compare it with what the agent said, and issue a verified-trust passport per run. 19 checks, 18 deterministic.Built · in production

Testbed

Agent 007 — public real-work benchmarks and a leaderboard where external agents are scored without instrumentation.Live since March 2026

Independent validation

Edinburgh Napier University assessed the platform, including Node Resolution, at UKRI TRL 4 — component validation in a controlled environment. Testing November 2025 – April 2026; letter ref. IV-XPLORE-2026, 20 July 2026. All test data synthetic with known ground truth; build frozen December 2025.

Scored cases

5,078

Same task, three architectures

Heading-level accuracy on HS classification: text-only ML 45.2 % (1,860 cases) → graph-augmented reasoning 90.9 % (564) → full pipeline under Node Resolution 100.0 % (540). Chapter-level: 61.8 % → 92.2 % → 99.8 %. No partials, failures or execution errors in the full pipeline.

Adversarial

500 cases, 11 categories. Malicious reaction rate 0 %. Executor reliability 1.000. Prompt injection 45/45; misinformation 45/45. Exact risk-level match 89.4 %; within one level 95.6 %.

Other suites

Risk-connection discovery 550 cases, 99.6 % count accuracy · sanctions screening 500 cases, 95.3 % status accuracy, 100 % on confirmed matches.

Not claimed

Explainability and the Data Integrity metric were outside this cycle — they are the objectives of TRL 5. An out-of-index stress test passed 41/80 and is recorded as a remediation target.

In production

First enterprise contract

Above £1M, multi-year. Global medical-device company; cardiac-screening agent inside a Holter product, 52 patients × 24 h ECG. Operated by the customer’s own team. Name under NDA.

Measured improvement

Weighted score 0.15 → 0.91 over thirty training iterations, without retraining the model — changes to prompts, tools, routing and policies only.

Verification on a foreign runtime

A third-party research agent verified through its own log, untouched: 79 % verified trust (92 % raw × 0.85 capture cap). 19 dossiers, 398 pages, 1,018 sources on a verifiable board.

Digital twins in operation

Three on one engine — customs (TRL 4), pharmaceutical logistics (built; 224 nodes, 9 types, 12 connectors), cardiac screening (production).

Public testbed

Since

March 2026

Agents

96 external agents · 80 runs scored on nine cases · five model providers. No agent modified, none given the ground truth.

Finding

On the Logistic Shocks case the harness moves the score as much as the model: same weights, five harnesses, 17 points of spread. Under cost caps agents stop citing sources before they stop answering.

Try it

Helpdesk Arena — five tickets in five minutes, humans and agents on one leaderboard.

Intellectual property

Patent

UK patent application on the Node Resolution method, priority September 2025. Further filings in 2026 on the orchestration and control loop and on verification of agent executions.

Academic programme

Done

TRL 4 assessment by Edinburgh Napier University, July 2026.

Submitted

TRACE-AI — a contradiction-aware provenance benchmark for agentic pipelines. Bid led by Edinburgh Napier with University College London; Xplore as industry partner. Decision pending.

In motion

Edinburgh Napier has been asked to coordinate a Scottish consortium on agent assurance; further universities in discussion. Concept note August 2026; scoping workshop October 2026.

Principle

Universities own the method — research questions, calibration, error bounds, the signature database. Xplore provides the instrument — capture, the deterministic engine, the testbed, live reference traces.

Recognition and membership

Cambridge Tech Week 2026

Live Pitch Competition — one of five finalists. Final: Wednesday 16 September, Cambridge Corn Exchange.

London Tech Week 2026

OneToWin finalist — top ten tech startups.

Tech Nation 2026

Breakout 50 — the UK’s most promising tech startups.

Vestbee 2026

Most innovative deep-tech startup, EU.

Member

Scotland’s Critical Technologies Supercluster.

Supported by

Interface Scotland.

The problem, in other people’s numbers

40 % of enterprise applications

will ship with task-specific AI agents by end-2026, up from under 5 % in 2025; by 2029 at least half of knowledge workers will work with, govern or create agents.Gartner, 26 Aug 2025

Over 40 % of agentic AI projects

will be cancelled by end-2027 — escalating costs, unclear value, or inadequate risk controls.Gartner, 25 Jun 2025

48 % ship without evaluation

and 74 % evaluate by hand, on a sample.LangChain 2026 (n = 1,340) · Chanl 2026 (n = 306)

Clinical agents fail silently on patient identity

Six models, 1.2 million tool calls: agents copied codes into tampered charts; detection of subtle identity faults near zero. “The central risk is misbinding, not miscoding.”Icahn School of Medicine at Mount Sinai, Int. J. Medical Informatics, Sep 2026

Quotable lines

Attributable to Xplore Intelligence or to the presenting founder; check the name with us before publication.

• “There is a gap between what an agent says it did and what it actually did. Everything we build closes that gap.”
• “Every agent vendor sells trust in its own product. Nobody sells the independent check. We are the roadworthiness test for agents.”
• “A model can be talked into anything. A graph cannot.”
• “The share of work an organisation can hand to agents is set by how much of it it can verify. We move that boundary.”
• “Universities own the method. We own the instrument. That is the only way a verdict becomes independent.”

Boilerplate

Xplore Intelligence is an AI research lab in Edinburgh building the infrastructure for governing and assuring AI agents. Its platform records what an agent actually did, checks what it claimed against that record, and certifies the result — for any agent, on any runtime, without modifying it. The method, Node Resolution, is the subject of a UK patent application with 2025 priority and was independently validated at TRL 4 by Edinburgh Napier University in 2026. Xplore’s first enterprise deployment runs inside a regulated medical-technology product; its public testbed, Agent 007, has scored 96 external agents since March 2026. The company is a member of Scotland’s Critical Technologies Supercluster and a Tech Nation Breakout 50 company. xploreintelligence.co.uk