How should architects handle Veras 5.2's guessed geometry?
Treat every question-marked element as an open design assumption, not finished BIM. Before native baking, capture the source views and the agent's stated guess. After baking, keep that status visible in Revit, Rhino or SketchUp until a named reviewer verifies geometry, type, dimensions and project intent.
Veras 5.2 puts a small question mark beside geometry the agent had to invent. That may be the most professionally useful pixel in the whole release. The model can be handsome. The question mark admits where the evidence ran out.
The official EvolveLAB release notes, published 24 September, say Agent mode can take a rendering, photo or sketch and build a 3D model step by step. It compares its result with up to 12 numbered source images and corrects mismatches. When it assumes something not shown, the affected element gets a question-mark marker with an explanation on hover. These are vendor claims; ArchiGen has not run the beta.
Why is a question mark better than a confident model?
Pictures are partial. A street elevation does not reveal the rear wall. An interior photograph does not prove the slab edge. A floor plan may show an opening while saying nothing about head height. Any system producing a complete model from incomplete views must infer.
That is not a software defect. Concealing the inference would be. The marker separates observed evidence from a plausible completion, a distinction architectural drawings already know through revision clouds, hold notes and “verify in field” tags. Agentic modeling needs the same candor, only closer to the object.
The danger arrives when visual fluency outruns evidential weight. A clean wall join looks decided. A family with a familiar name looks specified. Native geometry borrows authority from its host application, even when its dimensions came from the agent's best guess. ArchiGen's Astra scene acceptance test makes the same point at scene scale: inspect the model, not the flattering camera.
What does Veras 5.2 actually bake?
EvolveLAB says one Bake operation writes agent-made geometry into Revit, Rhino or SketchUp. In Revit, walls, floors and roofs become native elements; doors and windows use families; cameras become perspective views; some compound objects become families; and items without a native equivalent become Enscape proxies. SketchUp receives editable groups and Scenes. Rhino receives NURBS Breps and Named Views.
The release notes also say later bakes respect host edits unless the user chooses replacement. That is sensible behavior, but it creates two lines of authorship. One element may be agent-produced, then manually reshaped, then protected from a later bake. A screenshot of the result cannot explain that history.
| Status | What it means | Required action |
|---|---|---|
| Observed | Supported by a named source view | Check correspondence and scale |
| Assumed | Agent filled missing evidence | Retain marker and assign reviewer |
| Edited | Human changed baked geometry | Record author and reason |
| Verified | Accepted against project evidence | Close assumption with date |
Native is a file format condition, not a truth condition. A Revit wall can still be wrong. A Rhino Brep can still be a guess with excellent manners.
How can uncertainty survive the bake?
Do not rely on the Veras pane remaining open beside the model. The handoff needs a durable record that travels with the project file and its review process. Before baking, export or capture the element list showing every question mark, the hover explanation, the source image numbers and the scene identifier. Save that record with the model revision.
After baking, map each open assumption to something the host can keep visible. In Revit, that might be a dedicated parameter and review filter. In Rhino, use an assumption layer plus object user text. In SketchUp, assign a tag and attach notes in the issue tracker. The exact device matters less than the rule: uncertain geometry must remain queryable after it stops looking uncertain.
This extends the method in ArchiGen's client image assumption log. That log turns an attractive reference into questions about material, depth and buildability. Here, the questions attach to model elements before they quietly become coordination facts.
Can more source images remove the guessing?
Veras accepts up to 12 source images, according to the release notes. More views can reduce ambiguity, especially when they cover opposite sides and share recognizable control points. They do not produce omniscience. Images can conflict, use different design revisions or hide dimensions behind perspective.
Number the sources deliberately. Record origin, date and model revision for each one. Reject the tempting folder called “latest refs” containing three almost-latest facades. If two views disagree, the agent should not be asked to settle a design decision by statistical charm.
Use a coverage check before a run: which elevations are visible, which floors are represented, which dimensions are known, and which parts are absent? That short note gives the reviewer a map of likely invention. The agent's markers then confirm or extend it.
Which assumptions deserve the quickest review?
Review impact before polish. An invented chair may spoil a view. An invented opening can alter circulation, envelope area and coordination. An assumed roof profile can contaminate every downstream section. Give structure, egress, envelope and major dimensions priority even when the decorative errors are more photogenic.
Use the same severity logic as the image-to-3D acceptance gate: check scale, geometry, category, editability and provenance before admission. Agentic modeling broadens the object from a loose asset to a building scene, so the gate must become stricter, not friendlier.
ArchiGen's take: do not clean up the doubt
Software demos love completion. A blank scene becomes a furnished building while the cursor watches. Yet completion is exactly the wrong standard for a model reconstructed from partial evidence. Useful professional software should show where it knows, where it inferred and who accepted the inference.
Veras 5.2 deserves credit for putting guesses on the surface. Practices should resist the urge to clear those markers merely because the bake succeeded. Close each one with evidence or keep it open.
The model can lose a question mark when the project has earned the answer.
Evidence note: product capabilities were checked against EvolveLAB's official “Veras 5.2: Agentic Modeling with Astra & Opus” release notes, published 24 September 2026 and accessed 21 October 2026. The notes identify Agentic Modeling as beta and document source images, assumption markers and native bake behavior. ArchiGen did not install or run Veras 5.2; the review protocol is editorial analysis.