Two error distributions, two inference tasks
Non-standard engineering drawings contain furniture, legends, raster artefacts and incomplete enclosure. Our FY2026 investigation found that jointly predicting room boundaries and room classes conflated two different failure modes: an obvious room type could have a poor boundary, while a well-segmented region could carry the wrong class. The pipeline therefore separates single-class room geometry from vision-language room typing. The geometric representation remains editable rather than being flattened into an annotated image.
From polygons to a spatial graph
The unified floor-plan model carries room polygons, types and areas. Adjacency follows shared boundaries within a tolerance; containment follows polygon nesting. Indoor/outdoor classification is part of the spatial representation. These relationships are derived from geometry rather than produced by a second learned topology detector. Device and circuit association consume this substrate: a device on a shared wall cannot always be assigned by point-in-polygon containment alone.
An end-to-end measurement, not a universal accuracy claim
In June 2026, the combined image-to-space-model path achieved matched-polygon mean intersection-over-union of 0.84 against a 0.70 target, on a held-out set of approximately 17 drawings containing about 270 annotated room instances. This measures overlap for matched polygons; it is neither 84% recognition accuracy nor a score for either individual model. The evaluation is sufficient to distinguish large changes in usability, but not small differences between closely spaced results.
The operating point is a product decision
At inference confidence 0.45 and polygon suppression 0.4, the June evaluation recorded precision 0.712 and recall 0.622 on that held-out basis. A later reduction in confidence threshold traded precision for recall because drawing a missed room from scratch costs an engineer more than deleting a spurious candidate. Moving along the precision–recall frontier is configuration tuning, not a gain in model discrimination. Metrics from different thresholds must not be combined into an apparent new operating point.
What the experiments ruled out
Tiled inference recovered local detail at the expense of fragmenting large rooms across tile boundaries, degrading the reassembled space model. Increasing backbone capacity did not produce a recall gain in the recorded comparison. Adding annotations of the same kind did not move the observed frontier; that comparison used different validation splits and cannot establish a universal data-scaling conclusion. The resulting research direction is explicit wall, door and window structure, alongside a larger evaluation base. That structural extraction layer remained an open problem at the FY2026 reporting boundary.
Expert correction closes the loop
Manager provides the annotation and correction surface. Reviewed polygons and device labels can return as weighted training examples. The engineering artefact is consequently not a one-shot prediction: it is a proposed spatial model, its expert corrections and the downstream relationships that depend on it. This makes correction cost and geometric consistency first-class evaluation concerns.