Helios Substrate · Specification v1

This specification defines the Helios substrate at a level independent of any implementation. It covers the typed ontology, the bitemporal evidence model, the provenance contracts, the eight reasoning layers, and the closed-loop learning protocol that connects them. It is intended for reviewers, auditors, and engineering teams who are evaluating the substrate against an institutional deployment.

The substrate compiles an ontology into a typed graph. Entities, relationships, attributes, and constraints are versioned. Schema versions are pinned to every fact, every query, every recommendation. The substrate supports partial ontologies as a first-class concept: when an entity-relation falls outside the declared schema, structural ignorance is recorded as a first-class quantity, not silently coerced into Bayesian noise.

Every fact in the substrate carries two timestamps. The first records when the fact was true in the world. The second records when the substrate learned about it. This separation allows the substrate to replay any prior state of the world and any prior state of its own knowledge. Audits survive leadership changes, regulator inquiries survive employee turnover.

Every signal binds to its source. Every recommendation carries the evidence and policy that produced it. Provenance is not a logging feature, it is a contract: a recommendation without complete provenance is rejected by the substrate before it is emitted. Provenance contracts are typed, machine-checkable, and survive serialisation.

The substrate composes reasoning across eight layers: logical, combinatorial, continuous, probabilistic, causal, behavioural, temporal, and governance. A single recommendation may rest on a logically provable regulatory constraint and a probabilistic posterior over outcomes. The substrate preserves the type signature of each guarantee through the composition. Auditors can replay the composition step by step.

Outcomes feed back into the substrate. The model learns from the world it shaped, with full attribution to the decisions that shaped it. The learning loop is governed: a recommendation that produces a measurable outcome contributes evidence; a recommendation that produces an unmeasurable outcome contributes a known unknown. The substrate gets sharper from experience, not anecdote.