Good Monday from our atelier. Week 40 ran from a railing norm to a tram depot, from a cloth model over the A1 to a librarian for your project folder. Read together, the twenty pieces say one thing: the rules you write down are what last, whether they sit in an Archicad template, a Grasshopper definition or a one-page decision sheet.

Top stories:

Nine centimetres over the A1. Heinz Isler found the Deitingen Süd shells by hanging cloth and turning it upside down. That is how 9 cm of concrete spans 31 m on three points. The shells have stood since 1968 and survived a planned demolition in 1999. This was our strongest piece of the week because it is a method you can still run: before your next curved roof leaves the office, check the drawn form against a hanging-chain inversion. Then go and see Isler’s models at the gta Archiv. Saturday’s essay on dynamic relaxation is the digital sequel. The 1965 algorithm behind Kangaroo, Daniel Piker’s solver, finds a shape by letting it move until the motion stops. Only the resting shape is physical, so ship every definition with a plain-language objective panel.

The planet’s heat ledger. Gergana Gyuleva and her co-authors at ETH Zurich and Leipzig used Earth’s measured heat budget to narrow the climate-model warming range. The narrower range sits higher, at about 3.25 °C by 2100 on current policies. For a Swiss office, that is a question about weather files. Know which file each project runs on, and keep one warmer stress-test case in every overheating check.

Mattertal: protect or relocate. MGBahn’s own figures set out the decision. Keeping the valley line protected would cost CHF 220 million, and moving it into rock for good would cost CHF 470 million. The lesson reaches well beyond railways. Put protection capex, relocation capex, availability days and design life on one page before anyone orders the next protection net.

A librarian for the project folder. Our build of Andrej Karpathy’s LLM wiki treats it as a filing discipline. The rules file is what makes it trustworthy, with its citation, conflict and UNKNOWN rules written like a BEP chapter. Ingest the last three sets of minutes, then audit every answer against its raw source.

Shewchuk’s orientation test. Jonathan Shewchuk’s adaptive predicates are fast almost always and exact when it matters. On a million-point tilted grid, only 3 of about 9.4 million calls needed full exact arithmetic. Turning that research into a predicate people can actually call was half the gift.

←TODAY: In one week, a 1968 shell, a 1859 street grid and a 2026 climate paper all came down to rules someone wrote down. →3012: Zurich’s oldest buildings will be the ones whose reasoning was kept next to their geometry. Fulcrum: Form outlives its makers only when its rule set is filed with it.

Signal vs. noise: Two vendor announcements came with large numbers. WeatherNext 3 claims a 50% gain in day-ahead rain forecasts, but no published baseline or error bar exists yet. Gemini 4 Argon brings long-document AI to about $2, but the hard part for a Swiss office is routing: which project data may go to which endpoint under the nDSG, and with what cost cap per task. The capabilities are real. The headline figures are claims until you test them on your own site and your own tender. Cerdà’s 1859 Barcelona rule set and the Bauhaus Vorkurs were the quieter, sturdier signal.

Hidden gem: TITLIS turning a 1980s antenna tower into a Swiss cross deserved more readers. At altitude, the most valuable parts of a site are often the ones already there: the rock anchorage, the steel frame and the tunnel. Behind any retained structure stand the technicians who anchored it decades ago and kept the records. Credit them by keeping an “existing structure” layer in your model that records each member’s source, age and survey status.

Hack: Put a cheap filter in front of every raw zero test in your office’s geometry scripts, so that only the truly ambiguous “which side?” calls go on to exact arithmetic. The bound below is Shewchuk’s stage-A error bound. It works the same in GhPython or ported to a C# component in Grasshopper:

def orient2d_fast(a, b, c):
    l = (b[0]-a[0])*(c[1]-a[1]); r = (b[1]-a[1])*(c[0]-a[0])
    det, bound = l - r, 3.33e-16 * (abs(l) + abs(r))
    return det if abs(det) > bound else None  # None: escalate to exact

Count how many calls return None on your own grids. That count shows how often your script has been quietly guessing until now.

Atelier: This week’s thread for a Büro is ownership. Someone has to own the weather files, the railing favourites, the Kangaroo objectives and the wiki’s rules file, or nobody does. Monday move: give each of those four items a named owner on one shared sheet.

Looking ahead: PAZ-Rhino runs on 6 October. Bring the Rhino task your team repeats every week and leave with a small, readable Grasshopper rule you can explain. Watch for Gemini 4 Argon access opening, and start logging forecast rain against measured rain now, so that you will have a month of data when the next model claim lands. Collect your team’s three most repeated tasks before Tuesday.