01
From invention record to patent work product.
A workspace for invention records, prior-art research, claim development, drawings and counsel review. Outputs retain their source evidence and review history.
Prior art · claims · evidenceIndependent AI systems laboratoryEngland / 2026
Intelligence,armoured.
We develop software and formal methods for invention analysis, inference selection and verifiable AI governance.
BAYES' RULE
p(θ | D) = p(D | θ)p(θ) / p(D)
AI capability is becoming cheaper.
Reliable selection, governance and verification are not.
SYSTEMS IN DEVELOPMENT
Four research programmes.
Product work spanning invention analysis, model selection, verifiable governance and agent orchestration.
02
Selecting models under real constraints.
Research into model selection across capability, cost, latency, policy, provenance and uncertainty.
Decision theory · Bayesian inference · optimisation03
Verifiable evidence for governed AI actions.
Research into machine-checkable evidence for governed AI actions and independent verification.
Formal reasoning · verification · provenance04
Traceable agentic work.
Research into traceable agent workflows, explicit human approval points and reproducible outcomes.
Orchestration · evidence · human reviewMATHEMATICAL FOUNDATIONS
Decisions under uncertainty.
Optimisation, Bayesian updating, information theory and graphs give uncertain decisions a form that can be inspected.
Choice under constraint
x* = arg max f(x) s.t. gⱼ(x) ≤ 0Optimisation under policy, risk, budget, capability and latency constraints.
Belief after evidence
p(θ | D) ∝ p(D | θ) · p(θ)Bayesian updating makes the relationship between prior belief, observed evidence and revised belief explicit.
The value of uncertainty
H(X) = − Σ p(x) log p(x)Entropy quantifies uncertainty and helps estimate the value of further evidence.
Reasoning has shape
G = (V, E)Graphs represent claims, evidence, support and conflict without collapsing the reasoning into a single score.
Commitment before revelation
c = H(m)Hash commitments support later verification without publishing the underlying material at commitment time.
No single optimum
x ∈ Pareto(F)Pareto frontiers expose trade-offs between quality, cost, speed, control and trust.
MARKET CONTEXT
Relevant scale.
Published indicators for patents, generative AI spending and the projected AI market.
Patent applications filed worldwide in 2024
Source / WIPOWorldwide generative AI spend forecast for 2025
Source / GartnerProjected global AI market by 2033
Source / UNCTADRESEARCH DIRECTIONS
Problems worth working on.
Questions from current work in learning, evidence, security and systems.
Counterfactual policy search
How should an agent explore when other agents adapt? Parallel strategic alternatives and scenario-aware regret give exploration a memory of context.
Confidence under uncertainty
When should a conversational agent act, ask or defer? A confidence model can distinguish familiar turns from unfamiliar trajectories.
Monitoring live AI
How can a probabilistic model be monitored after validation? Track faithfulness, drift and stability as the surrounding system changes.
Identity inside encrypted traffic
Can encrypted traffic identify the software that produced it without decryption? Handshake metadata offers a protocol-level fingerprint for analyst review.
Hyperbolic evidence spaces
Can hierarchical evidence live in a geometry that preserves parent and child relationships? Hyperbolic embeddings make structure visible in a compact space.
Momentum in event streams
Can deal momentum be modelled as mutual excitation? A two-dimensional Hawkes process treats seller touches and buyer responses as linked event streams.
PRIVATE BRIEFINGS / PILOTS / PARTNERSHIPS
Request a briefing Discuss a research problem.
We welcome conversations about invention systems, governed AI infrastructure and applied research.