Foundation engineOntology Auto-Tuner
Refining knowledge graphs until they can carry weight.
A knowledge graph is only as useful as its consistency, and at real-world scale consistency does not survive contact with real data unless something enforces it. The Ontology Auto-Tuner is an automated refinement system for OWL ontologies and SHACL constraint shapes that detects structural weaknesses, proposes corrections and validates the result.
The Auto-Tuner is the common ancestor of everything else on this page. It is what allows a supply chain, a sensor estate or an evidence base to be represented formally enough that inference over it can be trusted, and it is available both as background IP within consulting engagements and as the basis of licensed applications.
- 1.4Mtriples demonstrated
- OWL and SHACLopen W3C standards
Early-warning assuranceSignal-readiness assessment
Knowing whether a warning deserves to be believed.
An early-warning pathway is only as useful as the evidence that it would sound when it mattered and stay silent when it did not. Published methods are usually assessed by their authors in terms of skill, whereas the decision-maker's question is narrower and harder, since it concerns whether a particular signal, on particular data and under stated uncertainty, is ready to be acted upon. Goodwood House provides that assessment as an independent layer in which observations, model outputs and forcing scenarios are bound into a provenance-carrying evidence graph, each warning pathway is tested against pre-registered positive controls, negative controls and stationary nulls, and the outcome is a traceable readiness judgement rather than a single headline number.
The layer assesses the detectors it evaluates rather than competing with them, so it can be applied to physics-based and statistical pathways alike. It draws on the Ontology Auto-Tuner for integration and on PRISM for claim-level provenance, and its worked example is CascadeWatch, whose full evaluation was frozen as a hash-verified audit package that a third party can reproduce figure by figure.
- Pre-registeredprotocols locked before data access
- Hash-verifiedfrozen, reproducible audit package
Supply-chain intelligenceDependMap
Finding the supplier you did not know you depended on.
DependMap maps multi-tier supply-chain dependency risk. A bill of materials is ingested, the supplier network is traversed beyond the first tier, and exposures that are invisible in any single procurement record become explicit, the canonical example being indirect rare-earth exposure that surfaces only at the second and third tier of a Western manufacturer's network.
The tool progressed through the Topcoder and Wazoku "Innovation Builders, AI Disruption" challenge, winning Stages 1 and 2 and reaching the Stage 3 vendor showcase. The working system comprises a live risk dashboard, an event feed and a manual simulation environment, exercised against a synthetic automotive dataset.
- Stage 3vendor showcase reached
Visit dependmap.com
Provenance and verificationPRISM
Provenance-assured verification for evidence you have to defend.
PRISM is a data model and enforcement layer for organisations that must know not only what their data says but where every statement came from and how well grounded it is. Built on the W3C PROV-O and Web Annotation standards with SHACL enforcement, it scores material along two independent axes, the assurance of the source and the grounding of the claim, so that downstream reasoning can be weighted accordingly.
PRISM was shortlisted to the final three in a UK government intelligence and security challenge, and the evaluators' feedback has been folded back into the design. It represents the practice's answer to a question that generative AI has made urgent, namely how an organisation reasons over material it did not author and cannot fully trust.
- Final 3UK government security challenge
Climate early warningCascadeWatch
A climate detector tested to its limits, and what survived.
CascadeWatch asked whether destabilisation of the Atlantic overturning circulation and the subpolar gyre could be detected in advance from sea-surface data using a network statistic, the leading eigenvalue of a correlation matrix across physically distinct ocean regions, rather than the single-site indicators that dominate the literature. Every analysis was pre-registered before data was touched, with the predicted outcomes and their interpretations locked in advance.
The observational record showed a multidecadal strengthening of coupling across the Atlantic network that passed its pre-registered tests. Model-world testing, which is the only setting in which a warning can be checked against a known collapse, did not support the method, however. Under full uncertainty the network signal could not distinguish the approach to a tipping point from a forced retreat that carried no such risk, and on the HadGEM3 freshwater-hosing experiments its trend was not significant against a stationary null on any run, whereas conventional temporal variance did warn of the collapse. The detector was therefore closed rather than kept alive for appearances.
What survived is the discipline. The protocol that exposed the method's limits now underpins the practice's signal-readiness assessment for early-warning pathways more generally. The work was presented at an invited seminar to Professor Tim Lenton's tipping-points group at the University of Exeter in September 2026.
Eradication inferenceEradication-confidence engine and planning tool
Proof of absence, and the difference between clearing a problem and merely holding it down.
The hardest question in any eradication campaign is not how to remove the target but how to know, defensibly, that removal is complete. Absence of detection is not detection of absence, and a campaign that cannot tell the difference either stops too early or pays for suppression indefinitely.
Goodwood House approaches this as an inference task. A stage-structured population model is combined with multi-modal detection data in a provenance-aware framework that generalises published inference-of-absence methods, and the output is a running, quantified confidence that eradication has been achieved, which separates a genuine path to zero from permanent suppression. The approach was developed against the invasive brown tree snake problem on Guam and validated in simulation against published field data.
The same thinking was then applied to an existing public-domain planning model for the problem. Working entirely from the published release, the practice reproduced that model independently from its equations, validated the reproduction against the model's own worked example, which it matched exactly on cost, coverage and the principal outcome probabilities, and in doing so identified where the original's results depended on the software evaluating it rather than on the science. The locked single-file model was rebuilt as a transparent, platform-independent planning tool that runs in a browser with no specialist software, and was then extended, again using only published data, so that it makes a decision-maker's central risk visible, namely whether a plan is on a path to permanent clearance or is settling at a plateau where every animal removed is replaced from the surrounding landscape. A published behavioural dataset was incorporated so that the physical placement of control tools informs the forecast rather than being left to judgement.
- Validatedagainst the public model's own worked example
- Public datapublished releases throughout
Open the planning tool
Formal trust verificationPhysical trust grammar
Machine-checkable arguments about physical security.
Security claims about physical systems, such as that a device is unclonable, that tampering is evident or that a measurement is fresh, are usually made in prose and evaluated by intuition. This work develops a substrate-independent formal grammar of physical trust primitives with an accompanying checker, allowing such claims to be composed and machine-verified rather than asserted. The work was developed in engagement with ARIA's "Trust Everything, Everywhere" opportunity space.
Government and securityProgramme work
Alongside the named tools, Goodwood House works within UK government innovation programmes on problems including the visual comprehension of complex sensor estates and thermal management for compact electronics, and has taken proposals through HMGCC Co-Creation and ARIA processes. The details of live programme work are not published here.