OmniCore v19. OmniCore v19: Architecture of a Reference Implementation for Coordination without Trust

The platform paper Coordination without Trust specifies a procedure for assembling coherent knowledge from mutually disagreeing intelligences, defined by four properties — traceability, confidence calibration, openness to attack, and resolution without merger — over a rank axis of the DIKWT type. That paper refers repeatedly to OmniCore v19 as an existing reference implementation, but no readable description of v19 has been published, so the reference resolves to nothing. This paper supplies that description. Two genres are deliberately kept apart. The platform paper describes an interface separable from any implementation; its value is that the procedure can be applied without knowing anything about OmniCore. The present paper describes one concrete architecture — a five-layer substrate L0–L4 divided by a domain boundary, a consortium of seventeen structurally distinct agents, a PostgreSQL realisation, and an external-review pipeline — and its criterion is therefore reproducible description rather than separability. The system is reported with an explicit boundary between what is designed and what is implemented, and the boundary does not fall where a reader might expect. The storage, promotion and protocol layers of the L0–L4 stack are implemented and deployed: the gold semantics of L0 (promotion under mandatory reviewer identification, audit journal, node history, and a generated coordination measure φ on nodes) exist in the omnicore_l0 schema, and the layer logic L1–L4 is realised as the Consortium Review Protocol (CRP) in omnicore_l4. What remains designed but not implemented is narrower and is stated precisely: the computation of Φ from a dynamical simulation at intervention nodes. The stored φ is an expert annotation, not a simulation result; the system does not at present compute its own coordination measure, and no claim in this paper depends on its doing so. The composition of the consortium is not asserted by design preference either. It was formed as a hypothesis from a full sixteen-agent run over a single work, and the resulting agent profile — which agent fires on which defect class, which is reliably silent, which duplicates an invariant — is recorded in a seeded corpus that any reader can reload and inspect. Two decisions taken from that run are expressed in the database schema rather than in prose, which makes them checkable against future data; §13 states what a single case can and cannot establish.