The Croatian industrial-automation shop whose production-system codebase the SYMPHONY substrate has to survive contact with. Demonstrator partner; supplier of half the held-out task instances for the O4 evaluation.
UP Robotics is the engineering shop that supplies SYMPHONY’s industrial-automation demonstrator codebase. Where Newcastle contributes the mathematical primary source, CREATE the architectural primary source, and Real AI the integration and coordination, UP Robotics contributes the production system that the substrate has to survive contact with.
Industrial code does not behave like open-source code. It is structured by IEC 61131-3 conventions, written in vendor-specific dialects (KRL, KAREL, RAPID, URScript, ladder logic, structured text), versioned across decades, and lived in by a small number of maintenance engineers whose institutional memory is the system’s rationale layer.
Ladder diagrams, structured text and function block diagrams from the programmable-logic controllers that run the line. Real-time-predictable, rarely commented, decades of authorship.
Vendor-specific motion routines, path libraries, calibration files for the manipulators and end-effectors. Often touched by many engineers over a single system's life.
Supervisory-control configurations, HMI screens, alarm databases, historian tags. The layer operators see — the one that survives engineer turnover.
Years of maintenance-engineer log entries — fault descriptions, repair notes, replacement-part references. The closest thing the system has to a rationale layer.
A neuromimetic knowledge substrate that works on a curated open- source benchmark but breaks on production industrial code is a publication, not a contribution to European industrial competitiveness. The O4 benchmark mixes 100 instances from OSS issue trackers with 100 instances from UP Robotics’s maintenance logs precisely to surface this failure mode if it exists.
200 engineering-task instances. Half sourced from open-source issue trackers, half from UP Robotics’ maintenance logs. Threshold: ≥ 20 % relative F1 improvement on task-relevant- subgraph recovery and ≥ 15 % expert-rated actionability, averaged across task classes — against three named baselines (frontier LLM agent, EAKG static-analysis pipeline, LLM + RAG).
If improvement holds on OSS but fails on industrial code, the §1.3 alternative path narrows the claimed scope and documents the domain-transfer gap as a scientific finding. The fact that this alternative exists at all is what makes the protocol honest.
UP Robotics and Real AI have agreed in principle on a joint venture as the primary exploitation vehicle for the SYMPHONY substrate. UP Robotics brings the customer relationships in industrial automation across Croatia and the wider region; Real AI brings the foundation-model engineering and the coordinator role. Read the full go-to-market and IP plan in the proposal.