Constraint-first architecture.
This is a long-form note on a shift the studio made during the Constraint Canvas commission (file IV.b) and that has since become a generalised principle in our work. We train into the constraints, not around them. The note is in part a written-down lesson and in part an invitation to argue with the position.
How most generative product work is built in 2025.
The default pipeline for generative product work at the time of writing has three stages. Stage one: take a foundation model. Stage two: fine-tune it on whatever client-specific data is available. Stage three: stack post-hoc filters on top of the inference output to remove generated artefacts that violate manufacturing, brand, regulatory, or aesthetic constraints. The pipeline is widely taught, widely deployed, and widely defended — it is the path of least resistance because every stage maps onto an existing toolchain.
It also, in our experience, almost always fails on bespoke work. The note is about why.
Three failure modes the standard pipeline carries with it.
1. The filter pollutes the workflow. Post-hoc filters reject a fraction of every batch of generated candidates. On a generic application, a 30% rejection rate is a minor inconvenience — the user sees the survivors and is happy. On bespoke work, the rejected candidates are not noise; they are the most aesthetically expressive part of the output space the model has learnt, and they are the candidates a designer would want to see. The filter selects against the very work the studio was hired to encourage.
2. The filter mis-trains the designer. If a generative workbench produces a hundred candidates a week and the filter accepts seventy of them, the designer internalises the seventy as the language of the work. Over months, the designer's drawn-by-hand work drifts toward the foundation model's distribution — not the client's. The filter is a slow contamination of the designer's eye.
3. The filter is a separate engineering surface. Filters drift. Models update. Rules change. A pipeline that depends on the filter behaving correctly is a pipeline that needs an engineer on call. We have walked into client studios two years after a typical generative deployment and found filters that had silently degraded to permitting near-anything because the threshold had been tuned during a deadline week and never restored. The filter was nominally working; the bespoke property was no longer present.
What we do instead.
The inversion is, written down, almost obvious: filter the training data, not the inference output. If every example in the training set respects the constraints, the model learns a distribution that does not contain unconstrained outputs. There is nothing to filter at inference because the model has never had unconstrained mass to begin with.
This is a deceptively simple statement and the consequence is large. It requires the studio to take the constraints seriously upstream of any modelling decision. The constraint sheet (see the companion note on the survey-room intake) is no longer a wishlist for downstream filtering; it is a specification of which training examples are valid corpus members. The studio's job in week one is therefore not to start training; it is to walk the corpus against the constraint sheet and to discard the examples that do not belong.
What we have learnt to look for.
1. Geometric, not aesthetic. A constraint sheet describes shapes, mass envelopes, kerning rules, plan-to-section ratios — things that can be evaluated mechanically on a training example. Aesthetic constraints ("our brand is honest", "the marks should feel modern") are valuable but they do not belong on the constraint sheet; they belong in the centre statement (see the bespoke-not-prompted note). The constraint sheet has to be machine-readable.
2. Authored by the practitioner, not the studio. The geometric rules of a foundry, a foundry-and-foundry, a sand-cast, a publisher's typesetting house, are knowledge the practitioner has accumulated over years and the studio has spent at most weeks with. We facilitate the writing. We do not invent the content. A constraint sheet authored by us against a practice we had visited twice would be a fiction.
3. Versioned, and dated. The constraint sheet drifts as the practice drifts. We version it explicitly — Rev 01, Rev 02, with the dates — and we re-walk the corpus against each revision. The drift is itself a piece of insight; we have on three occasions noticed a constraint silently weakening across years of practice and brought the noticing back into the studio's quarterly review.
The Constraint Canvas file, as a worked example.
The Constraint Canvas commission (work file IV.b) was the first time the studio applied the inversion end-to-end. The client — a Sheffield kitchenware foundry — produced an eighteen-rule constraint sheet over three afternoons with their chief moulder. We then walked their archive of 4 800 historical castings against the constraint sheet; 380 castings failed at least one rule and were excluded from training. The remaining 4 420 became the training corpus.
The resulting 220M-parameter voxel-diffusion model produces forms that respect the eighteen rules by construction. We have evaluated 9 000 generated candidates against the constraint sheet in a post-hoc audit (purely as a sanity check, not as a filter); the rule violation rate is below 0.1%, and the few violations are at the boundary of the constraints (e.g. a release angle of 1.49° against a 1.5° minimum), well inside the foundry's pour tolerance. The chief moulder's review remains in the loop on every commit.
The standard pipeline applied to the same brief in 2023 (we know, because we ran it as the studio's previous engagement model) had a 24% rejection rate, and the chief moulder had to spend two hours a week tuning the filter thresholds against a moving complaint list from the designers. The work is incomparable.
An honest accounting of the trade.
Constraint-first architecture trades expressive range for safety. The model cannot reproduce surfaces that violate the constraint sheet because those surfaces are not in the training distribution. Equivalently, the model cannot discover a form that breaks the existing manufacturing envelope — even when breaking the envelope might be desirable.
We have discussed this trade with every client on whom the studio has applied the inversion. In every case the client's position has been: the role of a generative workbench is to amplify the practice's existing voice, not to argue with the practice's process. Extending the envelope is the work of a chief moulder, a master compositor, a senior partner — people, not models. We agree, and we have not yet met a client who disagrees once the trade is named.
There is a corollary worth writing down: a constraint-first model becomes obsolete when the practice's constraints change. A foundry that retools to a different casting technology will need a new model. We consider this a feature, not a flaw — the model is a snapshot of the practice's geometry at a moment, and the snapshot ages with the practice.
What the studio holds, as of this filing.
Every generative system the studio has shipped since the Constraint Canvas commission has been built constraint-first. We expect every system we ship in 2026 to be built the same way. We will revise the position if we find a class of bespoke work the inversion fails on; we have not yet, and the principle has earned its place on the studio wall.
One specific invitation: if you have run the standard pipeline on a bespoke commission and you have either confirmed the failure modes here or escaped them in a way we have not seen, the studio would like to read your write-up. Notes that argue with our positions are the notes we learn from fastest, and we will print yours next to ours in the archive.