01
Description is not physical understanding
Modern language models are extraordinarily capable at describing beams, bearings, machines, buildings, and manufacturing processes. But a fluent description does not establish which physical object is present, where it is located, how it is oriented, which features belong to it, or what it connects to.
That difference matters in engineering. A system that says a beam normally supports a floor has demonstrated language knowledge. A system that identifies a specific member, measures its axis, establishes its connection points, and traces the evidence supporting those conclusions is beginning to demonstrate physical understanding.
Stratic Systems uses the term physical intelligence for this deeper layer: the ability of an AI-enabled system to maintain a structured, evidence-linked model of physical entities and use that model when explaining, comparing, inspecting, planning, or acting.
02
The substrate beneath reasoning
The central research hypothesis is that physical reasoning becomes more reliable when probabilistic models operate over a deterministic substrate rather than being asked to invent that substrate from language or pixels.
For engineered systems, that substrate can include object identity, geometry, canonical coordinate frames, finite axes, features, transforms, assembly hierarchy, adjacency, support, connection, provenance, and change over time.
Not every physical fact can be determined exactly. Vision may remain uncertain. Intent may require inference. Operational context may come from people or documents. The important architectural separation is that verified measurements, derived relationships, inferred meaning, and unresolved uncertainty remain distinguishable.
03
Why engineering is a useful proving ground
Engineering provides unusually strong conditions for testing physical intelligence because many aspects of the world are already represented mathematically. CAD models encode geometry. Assemblies encode structure. drawings and product data encode requirements. Inspection systems produce measurements. Machines and sensors produce state.
The challenge is not a total absence of information. It is that the information is fragmented across files, applications, coordinate systems, representations, organizations, and people.
NIST's work on the manufacturing digital thread describes the need to communicate structured product definitions across design, manufacturing, and quality while resolving differences between heterogeneous engineering representations. That is closely aligned with the infrastructure problem physical intelligence must solve before an AI system can reason safely across those domains.
04
A practical definition of progress
Physical intelligence should be evaluated by what the system can establish and preserve, not by how impressive the final sentence sounds.
A useful benchmark therefore starts with controlled questions. Can the system distinguish two geometrically similar members with different feature locations? Can it identify the local axis of a structural member? Can it preserve the source geometry that produced a measurement? Can it tell when a conclusion is unsupported? Can a field observation be associated with the correct engineering entity without losing provenance?
These tasks are intentionally less theatrical than general-purpose robotics demonstrations. They are also testable. Each capability can be compared against known geometry, known relationships, and known ground truth.
05
The Stratic Systems research direction
GeometryLab is being developed as the deterministic layer in this architecture. Its role is to transform CAD and engineering geometry into structured, queryable physical knowledge while retaining the evidence used to establish that knowledge.
SERA is the reasoning and orchestration layer. Its role is different: combine verified physical facts with language models, vision, tools, constraints, and goals while keeping inference separate from what has been measured.
The long-term progression is Observe → Understand → Reason → Act. The order is intentional. Autonomy should not begin with action. It should begin with a trustworthy representation of what the system believes exists and why.