Research
Fit and assembly order can be verified in software before a single part is cut.
A structural calculation proves the forces work. It does not prove that the parts touch where they should, that every joinery cut has a counterpart, or that part 47 can physically reach its seat past the 46 parts already in place. Kult checks all three on the same carved solids the CNC would cut.
Finding
Two families of geometric checks run on every generated house before machine files are exported: five mate checks (whether the parts actually fit together) and one insertability check (whether each part can be brought into place in the documented assembly order). Both operate on the posed, carved part solids: the same geometry the viewer renders and a cutting machine would execute. They therefore catch template and carving errors regardless of which template produced them, and in our own testing they have caught real errors that force-based checks are structurally blind to.
Method
The mate checks compute boolean intersections between neighbouring part solids and score five quantities, each against a pass limit: interior collision volume between two parts (mm³); end-support clearance from a member's end to its nearest mate (mm); bearing area as a share of the member's cross-section; orphan cuts (the share of a joinery cut left unfilled by its mate); and hole alignment (the share of a drilled hole's reach that lands in adjacent timber rather than in air).
The insertability check is a withdrawal sweep. Every part role carries a documented assembly order and an insertion direction. The check translates each part's solid outward along that direction in small steps and intersects each step against every part assembled earlier. Any overlap above a crumb tolerance means the part is trapped: it could not have entered without passing through something already in place. Insertion is blocked exactly when withdrawal is blocked, and withdrawal is the cheaper sweep to compute. This targets the most commonly documented field failure in comparable self-build systems: a part that cannot be inserted where the instructions say it goes.
The checks run across a regression corpus of 14 generated houses on every heavy test pass. They were additionally stress-tested in a review that pushed 30 hand-designed "reasonable family houses" through the full validation gate.
Numbers
| Item | Value |
|---|---|
| Geometric mate checks per house | 5 — collision, end support, bearing area, orphan cut, hole alignment |
| Assembly-order checks per house | 1 — withdrawal sweep per part against all earlier-order parts |
| Regression corpus | 14 generated houses, checked per part pair |
| Realism-review sample | 30 hand-designed family houses through the full gate (2026-07-05) |
| Example error caught | a screw pilot hole emitted 67.5 mm outside the timber for floor-level windows, a template bug invisible to force checks |
| Example error caught (visual origin) | gable studs floating one plate-thickness below the raked plate; a birdsmouth carved at mid-span where nothing bears |
Assumptions: parts are checked as rigid solids at nominal dimensions. Fastener kits and soft insulation batts are excluded because they have no meaningful solid-mate semantics. The checks are computationally heavy, so they run in the regression tier instead of on every interactive edit.
What this does not show
- It does not show that a physical build succeeds. The checks verify geometry at nominal dimensions, while real timber moves with moisture (several millimetres across a 220 mm face) and real cutting has tolerances. The physical proof ladder exists to test exactly that.
- The validation gate itself has had bugs. The 2026-07-05 realism review found the gate both over-strict (0 of 30 reasonable family houses passed as specified) and, in one case, under-strict, where a calculation error under-counted wind load on certain roofs. Both classes were found because the outputs are machine-checkable; fixes are tracked on the open board.
- No independent structural engineer has yet reviewed the engine's assumptions. That review is a named, scheduled step, and for any real build in Swedish safety class 2–3 it is a legal requirement.
Sources
- Kult validation modules — solid-mate checks and assembly-path check (internal, corpus-tested; the test suite runs the 14-house sweep per part pair).
- Kult internal review, "Validation-realism review" (2026-07-05) — 30-house sample, fail catalogue, rule classification.
- Kult internal research, "Mistake-proof joinery" (2026-07-03) — design-for-assembly evidence base; Boothroyd & Dewhurst DFA symmetry rules; Agrawala et al., sequenced one-part-per-step instruction research (Stanford).
- WikiHouse assembly guidance — orientation and sequenced instructions: wikihouse.cc/guides
Read next
One storm board cannot replace the racking layer
What happened when we swept a proposed one-board wall through the structural gate on 14 houses.
Five prototypes, each with a kill criterion
The physical proof ladder that cross-checks everything the software claims.
All papers
The research hub collects every published paper, and the note on why the work is open.