Building to 3D: From a Building Name to an Inspectable 3D Model

The system connects reasoning to executable scene state

An agent coordinates several tools around a dedicated project folder. Research tools retrieve descriptions, photographs and drawings; image inspection establishes what they actually show. An image generator produces material candidates. Blender MCP exposes scene inspection and Python execution, allowing the agent to build geometry, inspect views and make corrections. Saved Python scripts own the construction logic.

MCP carries tool requests and results. Blender’s Python API creates and changes scene data. The implementation protocol requires executing the same saved source through MCP, for example with runpy.run_path, so the live scene and a later rebuild share an implementation. Project-owned collections limit replacement to the reconstruction’s objects and preserve unrelated work. Separate background Blender processes handle expensive renders or verification.

The project.json manifest connects building identity, research, sections, textures, reviews and registered artifacts. A Python packaging helper checks these records and bundles the files. The offline report’s locally bundled Three.js viewer presents the GLB; the public exhibition is a subsequent presentation layer. These boundaries make a visible result traceable to its saved source and review history.

Research ends when coverage supports the exterior

A name is only an initial search key. Alternate names, address and target date help avoid researching a different building or combining incompatible historical states. The agent seeks complementary sources: official descriptions, architectural drawings, ground photographs, aerial views and detail images. It inspects their content, recording authorship, dates where known, retrieval status, observations, uncertainties and rights.

The published research acceptance gate assesses five categories: massing and scale, every distinct elevation, roof and courtyard relationships, materials, and defining features. A source count cannot establish coverage. Several copies of one frontal photograph still leave the rear unseen. A measured plan can support dimensions while a contemporary photograph establishes a changed roof or finish.

Scale anchors require interpretation. Plans retain their orientation, legends and graphic scales. Photographs carry perspective and occlusion; dimensions cannot be read directly from an uncalibrated view. Published dimensions, estimates and unresolved conflicts remain distinguishable in the assessment and modeling plan. Minor documented estimates can coexist with accepted coverage, while an unsupported roof structure or defining facade blocks it.

The default allowance is three total research rounds, including the initial search. Each round combines targeted queries, inspection and an immutable assessment. After insufficient coverage, the next round addresses specific gaps. At exhaustion, dependent modeling stops with recorded blockers and a request for useful additional evidence. Supplied resources can unblock the work without resetting consumed rounds. Research material remains evidence to interpret; external content does not become executable instructions.

Modeling proceeds from relationships to local detail

After acceptance, the plan fixes metres, axes, origin, geographic orientation and the target state. Sections follow the building’s architectural dependencies: shell, portico, apses, courtyards, roofs or tower as appropriate. Repeated bays and columns can share mesh data, while asymmetric features remain explicit. Evidence references explain the decisions that make this building recognizable.

The first model is a blockout. Clay elevations, overhead views and useful obliques test proportions, roof forms, connections, recesses and openings before ornament obscures them. The agent compares actual renders with inspected sources, corrects discrepancies and saves a reviewed checkpoint. Silhouette-changing features and architectural relief belong in geometry; fine grain belongs in materials.

The published Athenaeum modeling source makes this translation concrete. It builds explicit vertices and faces, revolves profiles for columns, and constructs rings, tubes and beams for architectural details. Face assignments select materials, UV coordinates encode surface scale, and repeated columns reference shared mesh data. Section ownership allows a local rebuild. Architectural interpretation therefore becomes inspectable parameters and construction operations: a reviewer can trace a profile or repeated bay to code, then compare its rendered result with the cited evidence.

Material generation targets neutral, front-facing swatches with no perspective or baked lighting. Exact prompts, available generator metadata, raw candidates and selection reasons are retained. Tiling and physical-scale checks examine repetition and grain at facade distance. Derived roughness or normal maps are authored estimates. A material-only pass must preserve geometry, transforms and UVs; generation unavailable under the current protocol leaves that phase blocked.

The default is two complete refinement sweeps, independently of the research allowance. Every planned section is visited in dependency order before another sweep begins. Stable cameras and lighting support comparisons; clay, close and grazing-light views reveal different defects. Each stage or section visit permits up to three corrections. Rejected attempts remain recorded, and a no-change review needs evidence. Failed gates block dependent stages; after each sweep, joins and the whole building are reviewed together.

Two buildings expose different failure modes

Romanian Athenaeum: evidence changed the rear roof

The Athenaeum implementation record separates an original exterior study from a later rear-and-roof correction. The baseline included three refinement sweeps but explicitly approximate concealed rear geometry. Three user-supplied aerial views subsequently resolved a semicircular rear roof, its straight inner chord, the circular connector and service equipment. Those private images are documented without being republished.

The correction protected the existing front and drum while revising five rear and dome sections through two sweeps. Recorded rejections included a projecting support wall, excessive equipment-support height, open roof wedges, six-fan units where the enlarged reference showed eight, and glazing in front of its mullions. These are concrete discrepancies with retained corrections. Further inspections still left concealed junctions and dimensions uncertain; that historical decision to proceed with disclosed estimates is distinct from the current skill’s research gate.

Material review revealed another boundary. A generated weathered-zinc swatch passed repeat and world-scale checks, yet the first portable browser review showed almost-black vertical zinc. A separate portable export softened its metal response while retaining the reviewed native Blender material. The saved records therefore distinguish geometric correction, material acceptance and export acceptance. A successful Blender render alone had not established browser appearance.

Sucevița: scale, openings and painted surfaces

The Sucevița implementation record concerns the Church of the Resurrection, excluding the surrounding monastery enclosure. Restoration drawings, a UNESCO dossier and photographs establish its triconch outline, rounded apses, unequal porches and tower. Graphic scales support an exterior length of approximately 39 metres. An official description’s 28-metre figure has unresolved measurement endpoints; the record preserves that conflict without assigning it a convenient interpretation.

Historical checkpoints contained filled openings and lateral UV banding. The accepted final history follows an explicitly authorized additional blockout correction, restored evaluated apertures and two complete sweeps through seven sections: plinth, body, apses, north porch, south porch, roof and tower. North paired arches and the south porch’s single large arches remain distinct. Opening and clearance checks address geometry that an attractive distant render could conceal.

Generated shingle, stone, plaster and timber textures provide estimated surface appearance. Licensed photographs supply recognizable fresco compositions, retaining attribution and the north derivative’s share-alike terms. Projection, pigment color and local placement remain approximate; a photographed hierarchy layout is reused across apses. Generated material imagery is never evidence of the real building, and inspecting a photograph for research does not itself authorize texture reuse.

Reproducibility requires several kinds of proof

The protocol defines five final checks: reopen the delivered .blend, verify packed textures, inspect evaluated geometry, check rebuild consistency, and review the GLB visually. Reopening in a separate Blender process tests the saved artifact rather than unsaved scene state. Geometry checks address relevant defects and opening clearances, with intentional sheets or overlapping trim documented instead of requiring universal watertightness.

A clean scripted rebuild compares stable geometry, transforms, UVs and material assignments. Repeating it in the same process checks that owned objects and resources do not accumulate. Byte equality between Blender files is not the criterion. The Sucevița record additionally retains relocated-rebuild evidence, addressing dependence on the original project location.

Its relocated-rebuild log records matching signatures, fifteen packed images and image paths redirected to relocated inputs. This supports the specific claim that the tested reconstruction could be rebuilt outside its authoring folder. The evidence remains bounded by that recorded environment. Changed source inputs or Blender versions would require another verification run before carrying the claim forward.

Export needs its own experiment. The agent builds an incomplete preview package, serves it locally and inspects the GLB from multiple angles, retaining screenshots before final packaging. Blender shader behavior can differ from portable PBR materials and viewer lighting. Simplifications and missing effects must be disclosed; native renders remain the material-appearance authority.

The manifest and packaging reference defines immutable reviewed artifacts, project-relative paths and matching SHA-256 hashes. Corrections receive new files and identifiers. Hashes establish file identity and integrity. The helper checks records, references, paths and registered files; it cannot establish resemblance, judge evidence sufficiency or prove that a claimed visual review occurred. Public records are selected, labeled snapshots, with some older checkpoints unavailable and no complete conversation transcript.

Six published models make the process inspectable

The Romania, in form. exhibition publishes the six reconstructions below. Each workflow record links its retained references, selected technical evidence, file lineage and limitations. The collection spans different architectural systems and research histories; it is a demonstration set rather than a controlled accuracy benchmark.

Assembly animations were authored after reconstruction from approved model exports. They illustrate architectural parts coming together, without asserting a historical construction sequence. Publication copies retime the retained twelve-second sequences to eight seconds while preserving geometry, textures, materials and transform samples. Recorded timing and video-decoding checks concern that presentation artifact, not the reconstruction’s geometric fidelity. Visitors explore prepared models; the website does not conduct a fresh reconstruction when a building is selected.