Generation is the last step; the constraints are the design.
Constraint Canvas is the work file for a four-month commission with a Sheffield kitchenware foundry that has been hand-casting brass and copper utensils since 1962. The client wanted a generative workbench that would help their two staff designers explore form variants for new product runs — without the workbench ever producing a form their existing foundry could not actually cast. The system shipped in November 2025 and has supported three product launches since.
A foundry that knew exactly what it could and could not do.
The client did not come to the studio looking for novelty. They came because they had run, over a decade, into a recurring friction in their design pipeline: the staff designers would produce form drawings on paper, the foundry floor would receive them, and roughly one in four would come back as unproducible — an undercut a sand-cast mould could not release, a wall thickness too thin for brass at their pour temperature, an internal radius the post-cast finishing process could not reach. The unproducible drawings were a constant, fixed-cost tax on the design week.
The brief was unusually clean. Help our designers explore form variants. Refuse to show them anything we cannot cast. Marisol Vegh took the file; Pita Selafyan handled operations; Eira Ashlin sat in on the foundry surveys because the manufacturing geometry was structural.
The inversion of the typical pipeline.
The default approach to a brief like this in 2025 was to fine-tune a foundation image-generator on the foundry's catalogue and to add post-hoc filters that rejected unproducible outputs. We rejected that approach twice in the survey week. The reasons we have written down in the work-file notes, and they have since become a studio principle:
1. Post-hoc filters are slow. If the model produces ten candidates and seven get filtered, the designer is being shown noise. The signal-to-noise ratio collapses below the threshold at which the workbench helps.
2. Post-hoc filters mis-train the designer. A designer who sees a hundred unfilterable forms over a month internalises the un-castable shapes as part of the language of what the foundry makes. The wrong vocabulary builds up. Six months later they will draw an undercut on paper because the workbench taught them undercuts were normal.
3. The constraints are the design. A casting foundry's distinctive output is shaped, at the level of form, by exactly the geometric rules the foundation model wanted to ignore. Removing the rules removes the foundry's voice. The brief was to amplify it.
Eighteen rules, all written by the foundry.
The first four weeks of the commission were not a model. They were eighteen rules, written by the foundry's chief moulder over three afternoons at the long bench, edited by Marisol and Eira in the survey room over two evenings, and signed off by the client in week four. The rules describe, in geometric terms, what the foundry can and cannot do. Minimum wall thickness at the pour temperature. Maximum unsupported overhang. Minimum internal radius for the post-cast finishing tools. Permitted angles for sand-cast release. The list is on the inside front cover of the Constraint Canvas binder; it lives in the studio in a folder marked "IV.b — Foundry rules — Rev 06".
The model never sees an un-castable shape.
The generative model itself is a 220M-parameter latent-diffusion model in a 3D voxel space, trained on a corpus of 4 800 of the foundry's historical product castings (digitised by structured-light scan, in-house, over week three of the commission). The training data was filtered through the eighteen rules before the model ever saw it. The model literally cannot reproduce the surface of an un-castable form, because the surface of an un-castable form is not present in its training distribution. There is no post-hoc filter on the inference side; there does not need to be.
This is the entire point of constraint-first architecture. The constraints are upstream of the model. The model is trained on a constrained distribution. The output respects the constraints because the output cannot do otherwise.
We are aware that this approach trades expressive range for safety. The Constraint Canvas cannot generate a form that breaks the foundry's casting envelope; it also cannot generate a form that extends the envelope into new territory. We discussed this with the client in week four; their position was that extension of the envelope was a job for the chief moulder, not for the workbench, and we agreed. The workbench is for the explored, not the explorable.
From canvas to mould pattern.
Every form the designer commits inside the canvas is exported, on commit, in two file formats: an STL mesh for visual review and an STEP file for the CNC pattern-making rig the foundry uses to produce sand-casting masters. The handshake between the workbench and the pattern-rig is automatic; the designer presses commit, the file lands on the foundry's pattern-rig server, and the chief moulder receives a notification with the geometry to inspect.
The handshake includes one human step we will not automate away: the chief moulder visually reviews every commit before pattern-making begins, and may reject a form for non-geometric reasons (aesthetic, brand fit, repair-ability after years of use). The workbench never bypasses the chief moulder. The studio considers this hand-off the most important integration in the work file.
Three product launches in six months.
Between November 2025 and April 2026 the foundry's two staff designers used Constraint Canvas to develop three product lines: a six-piece kitchen-tool set, a series of brass measuring beakers, and a set of copper candle holders for an exclusive retailer. The unproducible-drawing rate fell from roughly one in four (pre-canvas) to zero. We are sceptical of the zero; we expect it to drift up over time as the designers explore further from the centre of the corpus, and we have written that expectation into the maintenance schedule.
The canvas runs on a Threadripper-class workstation that the foundry owned for CAD work. Total roundtrip from designer-prompt to committed-pattern STEP file is under thirty seconds for any voxel envelope inside the casting volume. The whole installation fits inside the foundry's nightly backup; the kill switch is canvas vacate.
If your manufacturing envelope is the design.
Constraint Canvas is the corridor for clients whose distinctive output is shaped by what they cannot produce as much as by what they can. Foundries, ceramics studios, fine woodworking workshops, small fabricators with a strict house process. We do not take this corridor on for clients who want a generic product-ideation tool; the work depends on the existence of real, hard, geometric constraints that can be written down.