Engineering Intelligence: a constitutional architecture for AI over physical assets

This paper defines Engineering Intelligence as a distinct infrastructure category, states the three principles that constrain it, and specifies the seven-layer reference architecture and six physics-provenance states that follow from them.

No form, no email required. Cite and redistribute freely with attribution.

Industrial assets are extensively instrumented and poorly understood. A decade of data-first analytics produced accurate models and few changed maintenance plans, because the systems lacked a canonical engineering record, a non-bypassable physics constraint layer, and an auditable decision artefact. This paper argues that intelligence over physical assets must be constitutionally bounded by three principles — engineering defines the problem, AI assists the work, physics holds a veto over the result — and specifies the architecture those principles determine: a governed ontology, an engineering intelligence graph, a residual-publishing digital twin fabric, a deterministic unit-safe solver, a mandatory constraint gate, cited AI agents, and a signed engineering decision record. Six physics-provenance states are defined as a display requirement rather than as metadata.

Each layer depends only on the layer beneath it. No configuration permits layer 6 to bypass layer 5, or any layer to write to layer 1 without an ontology amendment.

How to cite this paper

ENGINPILOT (2026). Engineering Intelligence: a constitutional architecture for AI over physical assets. EP-WP-001, Revision 1.0. Bengaluru: ENGINPILOT Intelligence Systems.