System design for physical AI: one history per vehicle across two builders Engineer Muhammad Umar published an open reference architecture (CC BY 4.0, DOI 10.5281/zenodo.24048549) for unifying condition and maintenance data across two train builders for a Saudi passenger rail operator, built as a read-only projection over ten systems of record rather than a data lake or migration. The design uses bespoke adapters, Apache Kafka 4.4 in KRaft mode with per-source ordering and at-least-once deduplication, and a single write path into ConditionCase and AgentDraft objects, keeping the platform advisory and outside the safety case while fleet and passenger data stay on the operator's network. By Muhammad Umar. This is the engineering summary of an open reference architecture. The paper, its object model as JSON and the model register are free to reuse under CC BY 4.0: https://codeatoms.ai/passenger-rail-fleet-saudi-arabia/ https://codeatoms.ai/passenger-rail-fleet-saudi-arabia/ . DOI 10.5281/zenodo.24048549 https://doi.org/10.5281/zenodo.24048549 . The operator is described by class, never by name. A passenger rail operator in Saudi Arabia runs push pull trainsets of locomotives and coaches over long desert routes, with depots at the ends of the line and more trainsets coming from a second builder, maintenance included. A door fault seen on board three times this month is never joined to the work order closed last month for the same door, because the condition data arrives from the builder's platform in the builder's words and the depot closes work orders in its own. A locomotive whose traction converter is drifting is scheduled by the timetable rather than by its condition, and fails in the desert. When a train is late, the control room and the depot each tell a different story, and the availability figures the new contract will be measured against cannot be computed cleanly today. The design is a projection over the systems of record, never a copy of them and never a migration. A data lake aggregating the builders' feeds was set aside, because the question is one history per door with every fact tied to its source, and copied schemas cannot answer that with citations. The systems of record are read in place and never written; the single write path in the whole design is into the model's own two objects, the ConditionCase and the AgentDraft. The platform is advisory and non vital: it reads and drafts, touches nothing in signalling or train control, and so stays outside the safety case while staying inside the governance, with its introduction filed as a significance decision under the safety management system before any contract facing output goes live. Fleet and passenger data never leave the operator's network, and no third party model service is called. Ten sources feed the design. Eight are the operator's systems of record, all read only: the builder's remote monitoring platform, the on board diagnostic logs, the maintenance management system, the depot work orders, the timetable and train running system, the track condition monitoring service, the spare parts stock system and the passenger incident log. The other two are depot camera feeds, the inspection road and the siding overview, kept as context for an offered depot tier. Two adapter families carry them: integration adapters over the builders' documented APIs and the operator's own systems, and telemetry adapters for streams. Every fact enters stamped with its provenance class, a builder's diagnostic code enters exactly as the builder wrote it so translation happens in the ontology where it can be cited, and each source keeps its own clock offset, because every claim that this fault followed that closure depends on the builders' and the depot's clocks agreeing. No packaged integration product reconciles two builders' vocabularies, so the connectors are bespoke read paths. Apache Kafka 4.4 in KRaft mode runs inside the operator's network, ordered per source key so events from one vehicle keep their sequence, delivered at least once with deduplication on event identity so a repeat fault baseline never double counts. Connectivity from moving assets is intermittent, so the adapters buffer rather than block: when a trainset's link drops on a desert route, events queue and replay in order on reconnection. The time series store is chosen as a class and sized only after phase one counts events per day. Harbor holds every artifact, Keycloak holds identity on site, Prometheus 3.13 LTS with Grafana and Loki watches the running system, and Frigate NVR records the depot cameras. Time comes from a software time class over existing GNSS disciplined sources with per source offsets kept, because no requirement anywhere in the design is sub microsecond. The object model holds fourteen objects on Hyper Ontology v0.9. Three are assets: trainset, vehicle and serial numbered component. Three are events: diagnostic event, journey and delay. Six are records: fault with its confirmed cause, work order with its repeat flag, depot slot, diagram, incident, and the condition case the design itself creates. Part is a material, and the agent draft is a document the agents produce. One query walks the door: a diagnostic event is raised by a vehicle, the component is fitted on that vehicle and numbered as a part, the work order that closed last month is raised against the same vehicle, and the journey and delay on the trainset say whether the fault met passengers. Incident objects are minimized at ingestion to category and identifiers under the national data protection law, and availability is computed from journey and delay objects. All reasoning runs in one tier, the operator's existing data center in the country; there is no on vehicle compute and no edge tier in phase one, because telemetry already arrives through the builders' documented APIs. The work surface runs frontier class open weights, self hosted, on the 141 GB HBM class with eight GPUs per node, about 1,038 GB usable. The named candidate is GLM 5.3 at FP8: about 756 GB of weights fits one such node and leaves roughly 280 GB as the KV cache ceiling, which is why FP8 on one node was chosen over BF16 across two nodes or a larger 256 to 288 GB class node. That row is a candidate, not a commitment. No model in the register survived this run's live checks, so the paper asserts no licence terms and derives no GPU count until a person verifies a model row live. Nothing is trained, fine tuned or distilled. The node is ordered at engagement start, because its lead time into the country is the one dependency that can idle every workstream at once. Six services carry the work fleet ontology, condition watching, agents and memory, command center, sovereign platform, baseline and gates onto four surfaces: the command center's condition view, the diagram swap approval surface, the month end availability pack and the fleet engineer's approval surface. Every condition case carries an approving fleet engineer and a recommended depot slot, every agent draft carries an approval state and a citation set, and nothing reaches a depot slot or a diagram swap without a named person deciding. Every confirmed cause and accepted swap is kept as memory the next case reads, so the system improves by record, not by training, and a rehearsed export of the ontology and that memory is a phase one deliverable. Appendix A prices the node at about 570,000 US dollars to own over three years 542,000 to 594,000: server, support and power at the national commercial electricity tariff . Renting the same eight cards on the deepest three year commitment in AWS Middle East UAE is about 751,000, so ownership costs about three quarters of it, and on demand it is about 2.00 million, against which ownership costs a little over a quarter. Renting is a price reference only: AWS listed no Saudi region on the day the prices were read, and every rented option would carry diagnostic events, work orders and the decision record across the border. The rollout is two gated phases and 16 scope items 15, then 1 across six workstreams, traced against the 23 requirements in the requirement register. Phase one runs one route for one quarter against the operator's own baselines, written at signature; its exit gate is the fleet ontology of trainset and work order bound to a week of real extracts from both builders, with the joins proven before any agent is built. The headline measure is the traction converter drift score in sigma against a nominal of 1.0, and phase one ends with the simulated drift cased, swapped out of its diagram and repaired before it fails in the desert. Phase two onboards the new builder's vocabulary as a single item rather than a rebuild. Made with CodeNinja Praxis, the CodeNinja Atoms platform for designing physical AI systems. The object model imports into Hyper Ontology. Paper, ontology and register: https://codeatoms.ai/passenger-rail-fleet-saudi-arabia/ https://codeatoms.ai/passenger-rail-fleet-saudi-arabia/ One of the CodeNinja Atoms worked examples of physical AI system design https://codeatoms.ai/physical-ai-system-design/ : systems of record read and never written, a governed ontology joined by typed links, open weights on hardware the operator owns in its own country, and agents that draft while named people approve, the shape every physical AI https://codeatoms.ai/physical-ai/ design in the series is published in, with its object model as hyper-ontology/1 JSON under CC BY 4.0.