R&D
An in-house research programme with no client attached. It exists because the position we advise from — that probabilistic systems have to be constrained by a deterministic layer encoding the invariants of the domain — is a claim, and a claim has to be tested somewhere before it is charged for.
Structure before training
Our structural inference work derives relationships between entities from topological invariants rather than learning them by backpropagation. Nothing is trained, so nothing has to be trusted on the basis of a benchmark.
Determinism as a boundary condition
Where a model is used, it proposes and a deterministic layer holds the veto. Safety and correctness live in constraints on the output, not in instructions inside a prompt.
Commodity hardware as a constraint
Methods that only work on a research cluster do not reach a client's operations team. Our structural inference work resolves fifteen million rows in under two minutes on ordinary hardware, and that limit is a design input, not an afterthought.
Publish the negative results too
Research that only ever confirms the firm's existing position is advertising. What does not survive contact with a real system gets written up as well.
The papers, and where to read them.
Structural inference without training
A neuro-symbolic architecture in which the bonds between entities are defined from topological invariants rather than learned. The basis of the transaction-graph work that appears in client engagements as canonical sequence modelling.
Cognitive prosthetics and proxy interfaces
What happens to interface design when the user's agent, rather than the user, is the thing consuming the product — and what a system must hold fixed for that to be safe.
Research earns its keep when it changes a client system.
- Transactional platforms
- Canonical sequence modelling: one transaction structure specialised per vertical, rather than a separate model per business line.
- Catalogue and pricing systems
- Margin derived from procurement reality with market data setting position — structure first, then the model that operates inside it.
- Regulated catalogues
- Domain rules — dose, form, prescription status — encoded in the data model rather than enforced in the interface that happens to be in front of the user.
- Assessment and behavioural signal
- Instrument design for behaviour captured under load, scored against a structural model rather than a self-report scale. Shipped as NeuroFrame.
Intuition is not a strategy.
The through-line from the research to the client work is a single claim: probabilistic systems have to be constrained by a deterministic layer that encodes the invariants of the domain, and the value lives in the invariants rather than in the model. It is the same argument applied to a transaction graph, a pharmacy catalogue and a bilingual contract.
Invariant
Define what must remain true before deciding what should change.