TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Studio Practice · AGE 19—22 · FACTUAL · EVOLVING
Making, not just drawing
Tolerance stands between a generated form and a made thing, and an image has no ruler.
The idea
Between a generated form and a made thing stands tolerance. Materials move, tools have limited reach, joints need to fit, and a millimetre matters. No single file carries all of that. The chain runs: design geometry; fabrication constraints — tool, stock, kerf, tolerance; toolpaths from both; a simulation to check them; post-processing into one machine's instructions; then the object, measured. An image holds none of these, because an image has no ruler. Digital fabrication is where generative outputs meet the physical world, and the physical world does not negotiate. A render fails on a screen. A panel fails on site.
Why it matters
The last metre between file and object is where most of the knowledge lives. A student who can generate a pattern and cannot get it made has learnt the cheap half.
See it in the studio
A generated brick screen, beautifully lit. The mason asks the bond, the module, how the corner turns, what holds the perforated course. Nobody asked the image, because the image could not have answered. The same pattern laser-cut at model scale comes back with every slot a kerf too wide.
Watch for this
Treating a mesh exported from an image-to-3D tool as a fabrication file. It is a likeness of a surface. It has no thickness you chose, no tolerance, no knowledge of the material it will be cut from — and material behaviour is never "in the file": it enters as the feeds, speeds, offsets and allowances a person sets when the toolpaths are made, and it is checked on the bench.
Try it
Take one generated pattern. Cut or print a hundred-millimetre piece of it. Measure the deviation between the file and the object. Write down everything the file did not know that the machine needed — the kerf, the material thickness, the minimum feature the tool could cut.
Prove it
Explain where tolerance enters between file and object, and why an image model cannot hold it, with one measurement from your own test.
How it works
Robotic and CNC fabrication in architecture is a research field in its own right — ETH Zurich's Gramazio Kohler Research has spent two decades on it — and its central lesson is that the geometry and the making are designed together, with the machine's constraints inside the design from the start, and the toolpaths, the simulation and the post-processed machine code treated as stages of the design rather than an export at the end. Generative image models sit outside that loop: a 2025 study that tested three diffusion models on vernacular pigeon towers found they reproduced geometric pattern reliably and misread material and climatic reasoning — the look of a construction without its logic. Tools that feed fabrication constraints back into generation are active research, which is why this card is EVOLVING; the tolerance itself does not change.
What this idea builds on
What this idea opens up
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.