Verification infrastructure for high-consequence AI.
LogicalMetricsAI sits alongside existing AI systems as an independent verification layer. We decompose AI outputs into individual claims, check each one against version-controlled authoritative sources, and produce a structured audit trail before any decision reaches a human.
Built for healthcare, finance, insolvency, and regulated environments where unverified AI cannot be deployed.
Claim-level verification
Every AI output is decomposed into individual claims. Each claim is checked against version-controlled authoritative sources, including NHS guidance, regulatory statute, and domain-specific evidence libraries. Claims that match verified evidence pass through. Claims that cannot be verified are classified and returned as contradicted, unsupported, or uncertain alongside supporting evidence spans and safety analysis.
Deterministic safety overrides
Probabilistic AI outputs are checked against deterministic safety rules. Dosing limits, contraindications, statutory thresholds, and other hard constraints sit outside the model and cannot be bypassed by statistical inference. Safety rules are explicit, inspectable, and version-controlled.
Structured audit trail
Every claim, evidence span, confidence score, and safety flag is written to a structured JSONL ledger in real time. Decisions are traceable, defensible, and ready for regulatory, legal, or clinical scrutiny.
About
A different architecture for verifying AI
LogicalMetricsAI was founded on a single insight: AI systems struggle because probabilistic inference cannot reliably distinguish truth from falsehood in high-consequence environments.
Our approach separates verification from inference. Truth is maintained externally as a verified, version-controlled evidence library, while reasoning models are used to identify patterns of contradiction, unsupported assumptions, and reasoning failure.
Claims are first checked against authoritative evidence. Deterministic safety rules operate independently of model output and cannot be overridden by statistical inference.
This architecture forms the foundation of AARF, our Auditable AI Reasoning Framework. AARF is model-agnostic and operates alongside existing AI systems, making their outputs auditable, inspectable, and deployable in environments where trust is currently missing.
Reasoning error coverage
AARF includes specialist reasoning models trained to identify six categories of reasoning failure:
- Generalisation out of population (GOOP): applying findings from one population to another where they do not hold.
- Reversal of rules: inverting a rule's direction or polarity.
- False causation: treating correlation or sequence as cause.
- Missing reasoning step: skipping a required intermediate step.
- Misapplied rule: applying a rule outside its valid scope.
- Unsupported assumption: introducing a premise that is not grounded in evidence.
The models currently achieve between 89 and 96 percent accuracy on held-out synthetic test sets and operate alongside deterministic verification and evidence-grounding systems within the wider AARF pipeline.
How AARF works
AARF operates as a five-stage verification pipeline.
Input
A clinical document, legal case, financial workflow, or AI-generated output enters the system. The Router Agent identifies the domain and selects the appropriate verification path.
Retrieval
The Librarian Agent retrieves version-controlled authoritative evidence. Sources include the BNF, NICE, NHS England guidance, insolvency statute, and other domain-specific libraries.
Verification
Claims are checked against the authoritative evidence library using evidence comparison, contradiction analysis, and deterministic safety rules. Claims that cannot be verified are classified and returned with supporting evidence, confidence scoring, and structured failure analysis. Numerical and statutory constraints are enforced independently of model output.
Output
The clinician, practitioner, or decision-maker receives a structured report covering verified claims, contradictions, safety warnings, and confidence scoring. Final authority always remains with the human.
Audit
Every stage is written to a structured JSONL ledger. Claim text, evidence spans, confidence scores, safety flags, and final classifications are retained and inspectable.
Validation
Independently assessed and externally recognised
AARF has been tested across 15 synthetic NHS prescribing scenarios covering lethal dosing errors, contraindications, and dosing frequency conflicts. Detection rate on this synthetic test set was 100 percent with zero execution failures. Real-world clinical validation is the next stage of work.
Independent technical assessors from the Innovate UK Agentic AI Pioneers Prize scored LogicalMetricsAI 8.3 out of 10 on user and workflow fit, confirming the system maps credibly to real NHS clinical workflows.
LogicalMetricsAI is included in the Tech Nation Breakout 50, an annual list of emerging UK technology companies to watch.
Team
Andrew Jackson
CEO and Founder
Andrew is the architect of AARF and the originator of the company's core intellectual property. The founding insight behind AARF emerged from his research into misinformation detection and AI reasoning failure.
Richard Clarke
Systems Integration Director
Richard brings two decades of data analysis and machine learning experience across central government and multinational private sector organisations, specialising in cyber security and intelligence.
Gary Dobson
Commercial Director
Gary brings over 30 years of senior technology leadership experience, including time as Sales Director for a global technology company.
Talk to us about a pilot
If you are deploying AI into healthcare, finance, insolvency, or another regulated environment, we run paid technical proofs of concept that demonstrate AARF against your workflows, evidence sources, and decision pathways.
Ready to build the future? Let's start the conversation.
For more information, check out our Privacy Policy and Terms of Service. You can unsubscribe at any time.