The research behind understanding, reasoning and responsible action. Technical reports, lab notes and the questions we are still working to answer.
Evidence before claims. Learning before certainty.
Ideas that can be examined.
Machine competence does not have to live inside a model's parameters. A capable system can be split into a general policy that decides what to do and a library of specialized, inspectable, executable components that do it. For decisions tha…
Read paperMay 2026 · 32 pagesA bitemporal evidence substrate makes a world model 'living'. Each fact carries valid time (when it was true) and transaction time (when we believed it). An append-only ledger separates source of truth from a derived memory index. A continu…
Read paperApril 2026 · 28 pagesHigh-risk environments need ontologies that are typed, axiomatised, and checkable. We use language models to propose and SMT solvers to dispose: schema generation, semantic bridge axioms, and safety properties become satisfiability problems…
Read paperMarch 2026 · 26 pagesDiscrete agent choices block gradient flow in classical agent-based models. We relax via Gumbel-softmax, calibrate by simulated minimum distance, define interventions as program transformations, and treat counterfactuals as fixed-noise repl…
Read paperFebruary 2026 · 24 pagesΘΕΜΙΣ interposes a verification gate and an append-only, hash-linked, Ed25519-signed ledger between decision engines and consequential actions. A neurosymbolic firewall checks policies via SMT. A PII gateway enforces egress rules. Four secu…
Read paperFebruary 2026 · 18 pagesOn why governed, living world models, built inside European institutions, are a better foundation for organisational intelligence than autoregressive assistants alone. Five claims, defended.
Read paperFebruary 2026 · 42 pagesThe formal specification of the Helios substrate: typed ontology, bitemporal evidence, provenance contracts, and the closed-loop learning protocol that connects them. Implementation-independent.
Read paperJanuary 2026 · 24 pagesWhen should a decision system refuse to answer? We propose a calculus of provable refusal: certify that a query lies outside the model's evidence, and return a citable bound rather than a hedged extrapolation.
Read paperDecember 2025 · 12 pagesThe seven principles by which a Helios world model is built, governed, and held accountable. Our equivalent of a constitution: short enough to enforce, long enough to mean something.
Read paperNovember 2025 · 36 pagesHow a Helios world model maps to the obligations of the EU AI Act across risk classification, transparency, human oversight, robustness, and post-market monitoring. Auditor-ready.
Read paperOntologies now compile to a typed graph with version pins on every entity-relation. Reasoning runs always cite the schema version. Customers can roll a substrate back to any prior schema and replay decisions deterministically.
Internal preprint. We argue partial-ontology uncertainty is a first-class quantity, distinguishable from aleatoric and epistemic uncertainty over data, and composable across all eight reasoning layers without leakage.
We ran a six-week adversarial campaign against the governance layer of a banking substrate. The team probed ontology gaps, permission boundaries, and policy staleness. Eighteen findings, all reproducible. Mitigations land in v1.6.
We are formalising how a recommendation that rests on a logically provable regulatory constraint and a probabilistic posterior over outcomes can compose without losing the type signature of each guarantee.
Three new hires from European universities. A first quarter of joint engineering with a sovereign wealth fund and a tier-one bank. Long evenings, short emails. The work feels right.
A short note on why our lab focuses on organisational intelligence rather than autoregressive assistants. The world does not need a smarter answerer. It needs a substrate that remembers, governs, and learns from outcomes.
Helios Brain now records every fact with two timestamps: when it was true in the world, and when we learned about it. Audits can replay any prior state. Bitemporal is no longer a research property, it is a customer guarantee.
Customer ontologies are always partial. When the substrate encounters an entity outside its declared schema, current practice silently coerces it. We are formalising partial-ontology uncertainty as a first-class quantity.
Explore this question togetherMost decision systems extrapolate beyond their support and emit a softened warning. We are researching a refusal calculus where the substrate can certify a query lies outside its evidence, and return a citable bound.
Explore this question togetherA single recommendation may rest on a logically provable regulatory constraint AND a probabilistic posterior over outcomes. Today these compose unsoundly. We are exploring lifted probabilistic logic.
Explore this question togetherForecast-accuracy backtests reward the wrong objective. A decision system should be evaluated on the decisions it enabled and the regret it accrued, not the RMSE of intermediate predictions.
Explore this question togetherWhen the action space is high-dimensional and partially confounded, causal identification under do-calculus is rarely possible. We are researching practical relaxations: partial identification, sensitivity-bounded effects.
Explore this question togetherA world model that learns hourly, weekly and quarterly must detect distribution shift at each scale independently. Standard drift detectors collapse the scales.
Explore this question togetherDomain experts encode knowledge that data alone never captures. Turning that tacit knowledge into versioned, citable, governable evidence, without flattening it into brittle rules, remains an unsolved problem.
Explore this question togetherCompliance is the last line. We red-team it internally with multi-agent adversaries that probe rule boundaries, jailbreak permissions, exploit ontology gaps, and trace data exfiltration paths.
Explore this question togetherA question worth exploring together.
For research collaborators, technical contributors and people testing the limits of decision intelligence.