CORE CONCEPTS
Structured output
Rowe returns typed output objects for tasks such as fit checks, placement, stability, prediction, reconstruction, action, and optimization. Four Cad consumes reasoning_scene.v2 inside those outputs to render objects, relation lines, collision volumes, support patches, measurements, issue regions, fit results, comparisons, geometry changes, and grounded text.
Typed, not prose
Every task returns a typed object rather than a paragraph to parse. A fit check returns a verdict, a clearance number, and the entities responsible; it does not return a sentence describing them.
reasoning_scene.v2
Spatial outputs can carry a renderable reasoning layer: objects, relation lines, collision volumes, support patches, measurements, issue regions, comparisons, and geometry changes. Four Cad renders exactly this — there is no separate visualization format.
Traceability
Findings reference the geometry they were computed from. That is what makes a result reviewable rather than something to take on trust, and it is the reason we keep the structure instead of flattening it.
Response fields
Returned on every successful response.
idstringUnique response id.
objectstringThe object type. Usually response.
modelstringThe model that generated the response.
outputarrayTyped spatial outputs for the requested task. When the request carried a cad_scene, understand, describe, and compare add output[].data.cad_scene: named objects with absolute bounds and parents, connectors with direction, spatial relations, title block, and callout counts; compare adds output[].data.cad_scene_diff (added, removed, moved, re-parented, rewired, dimensions changed).
usageobjectInput spatial tokens, output spatial tokens, projected input cost, projected output cost, and total estimated cost.
request_idstringRequest id for support and debugging.