We investigate how AI systems reason, make decisions and demonstrate what they have done. Our research connects invention, governed inference, verifiable evidence and human oversight through working prototypes and testable ideas.
Science-led engineering: explicit models, testable hypotheses and evidence you can inspect.
Our work draws on decision theory, Bayesian inference, optimisation and formal reasoning. We make assumptions explicit, test behaviour and use evidence to challenge the model. The maths earns its place by explaining a real engineering decision.
x* = arg max U(x), x ∈ E(policy)OpenScorpion’s routing problem starts with an eligible set. Within it, utility makes quality, cost and latency trade-offs explicit. The weights express priorities; they need evaluation against actual outcomes.
p(θ | D) ∝ p(D | θ) p(θ)Bayesian reasoning makes the relationship between prior assumptions and new observations explicit. It informs how we think about uncertain claims and model performance; a probability still needs calibration and testing.
G = (V, E) · h = H(evidence)Graphs represent claims, evidence and conflicts; hashes bind a record to specific material. These ideas underpin our work on compliance and traceable delivery. A valid signature checks integrity, while the underlying evidence still needs scrutiny.
An answer is only the start.
The hard part is making it usable and accountable.
Each project tackles a different point where AI needs evidence, judgement and control. These are active development efforts; specific capabilities and readiness vary.
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 · evidence
View full-size We are building a routing layer between applications and model providers: apply policy first, compare eligible routes, and retain the decision evidence. The problem is balancing answer quality, spend and response time without losing control.
Policy → eligible models → route → decision recordWe are building policy evaluation and signed evidence bundles across an engine, platform, review UI and independent verifier. The aim is to preserve what was checked, which policy applied and who approved it.
Policy → evidence → decision → independent verification
View full-size Ariadne connects agent activity with repository work, tests, reviews and delivery evidence. The private console tracks connected Macs and VMs, including work before a PR exists. We are extending that visibility to cloud agents and bot workers.
Orchestration · evidence · human reviewFollow the change. Inspect the evidence. Keep the decision with a person.
Connect agent activity on Macs and VMs to the repository, even before a pull request exists.
Agent · workspace · branchBring code changes, pull requests, tests and reviews into the same delivery story.
Change · checks · reviewInspect the supporting evidence and outstanding checks before deciding the work is complete.
Evidence · human judgementA continuous thread of evidence. Activity is the starting point. Reviewable delivery is the goal.
A bit of fun alongside the main research: daily interactive experiments in maths, physics and computing. Explore an idea, adjust the parameters and see the behaviour for yourself.
No tile can see the whole maze. Each one only checks the four tiles touching it and says: I am one step further than my closest neighbour. A wave floods out, the maze rises into a staircase, and the way out is just 'walk downhill'. Smash it and it heals.
Explore the demoExplore how a small network changes its decision boundary as it trains.
Try the experiment INTERACTIVE DEMOChange the conditions and see how a flow develops.
Try the experiment INTERACTIVE DEMOExperiment with the relationship between motion and gravity.
Try the experimentAn informal series driven by curiosity. The latest experiment updates automatically from the published Interesting X collection.
We welcome researchers, engineers and collaborators working on invention systems, governed inference, verifiable decisions and the behaviour of AI agents.