Most AI floor-plan tools draw something plausible and hope you will not measure it. RONCHAMP takes the opposite route: a small in-house model proposes intent, a constraint solver builds the geometry, and an independent verifier decides whether the result may be called a plan at all.
The chain
- 1 · Archive retrieval. If a verified plan in the office archive matches the parcel and program closely, it is adapted first — real, approved layouts beat generation.
- 2 · Intent model. A small plan model trained only on our own approved archive (no scraped data) proposes the room adjacency intent in ~0.4 s on plain CPU.
- 3 · Constraint solver. A CP-SAT solver turns intent into geometry inside the buildable polygon: room bands, door reachability, stair placement, façade access.
- 4 · Independent verification. A separate process re-measures the result: overlaps, program fidelity, polygon integrity. Fail → the output is honestly labelled a study, not a plan.
- 5 · Evidence. Every candidate carries its decision timeline — you can inspect any stage and branch from it with one change.
Why this is different
Generative models are excellent at suggesting and terrible at guaranteeing. Solvers are the reverse. RONCHAMP splits the job accordingly: the model only proposes; hard rules (door access, minimum areas, overlap bans, setbacks) are enforced by the solver and re-checked by a verifier that does not share code with the generator. The result is auditable: the same seed reproduces the same plan, and every accepted plan feeds back into the archive that trains the next model.
Speed, honestly measured
Typical detached-house generation completes in under 10 seconds, including verification and evidence; multi-storey and large programs take longer and say so. Stage timings are visible per run — no vague 'a few seconds' claims.
See it on your parcel
Fetch a parcel, set the program (bedrooms, bathrooms, options) and generate — then open the proof card and the timeline. Create an account or read the zoning guide first.