The Week the Dependency Graph Came Due: Spare Paths, Open Files, and Who Resets It at 2am
Week 2026-W26 recap: Sakana Fugu's spare path, Neural CAD vs CADAM's openable files, and the single points of failure hiding in your AEC stack.
From the cartographer’s desk: the past seven days read like a single audit. Strip away the cover stories — AI CAD, soft robots, map projections — and the same question runs underneath all twenty pieces. What does this depend on, and what happens the day that one thing goes dark? This was a week about hidden topology, and the field finally started drawing it.
←TODAY: In 2026 a single LLM outage, a single PRNG, or a single proprietary file format can quietly become the load-bearing wall of your whole studio. →3012: Resilient inference becomes civic infrastructure — we stop trusting any pipeline with exactly one path through it. Fulcrum: Redundancy is only cheap if you draw the dependency graph before the outage, not during it.
Top stories:
The week’s spine was Sakana’s Fugu — an orchestration model that routes tasks across a swappable pool of frontier LLMs. The real product isn’t a higher benchmark; it’s the spare path in the graph, redundancy sold as an API. For any studio whose computational stack calls exactly one model by name, Fugu is the argument for a provider-agnostic abstraction layer this week, not after the storm. This is the single-point-of-failure beat in its purest form.
Running close behind, the Neural CAD thread asked the same question from the file end. Autodesk’s foundation models generate editable geometry from speech and sketch — a genuine gain at the front of the design funnel — but the burden downstream is unchanged: that geometry still has to survive the round-trip into a coordinated execution model. Adam’s CADAM answered with the responsible pattern: AI writes parametric OpenSCAD once, then humans tune sliders with zero further model calls, on text you can diff and open forever. Two tools, one acceptance test — when the vendor vanishes, can someone still open the file?
On home ground, two Zurich dates anchored the week: the Archicad-meets-KI Powersession (24 June) and the three-day KI Sommer Camp 2026. The framing held firm — KI earns its hour when it drafts and tags inside Archicad and leaves the Freigabe to you, and AI competence now means a clean local workflow that reads your project data without surrendering it to a vendor’s cloud. That is the same resilience instinct as Fugu, scaled down to one desk.
And for the practitioners who like their systems load-tested by physics, Jason Davies’ projection gallery and the catenary solver were the quiet standouts: every flat map lies, and the best designer of a load-bearing shape is the force itself. Both are reminders that the maths under a developable façade or a swisstopo CH1903+/LV95 georeference is a decision, not a checkbox.
Signal vs. noise: The genuine signal was infrastructural — Fugu’s swappable routing, CADAM’s diffable output, the three-class secrets scanner cutting false alarms 33%, and even a small C++17 random library exposing std::random_device as a quiet single point of failure in verification-grade calcs. The noise was the “robots replace humans” wave: Japan’s teleoperated construction giant is an operator-amplifying machine sold in autonomy’s vocabulary. Read it from the floor, not the press release.
Hidden gem: The blurry-mean trap (arXiv:2606.12987) deserved more attention than its score got. A model can ace SSIM, cosine similarity, or mean distance and still be wrong, because averaging metrics reward the safe blurry mean and hide whether the thing is controllable. Every AEC team commissioning a generative tool should write acceptance tests around a distribution metric (FID/KID) plus a controllability check — judge the tool on which distribution it matches, not how close it lands to one safe answer.
Hack: This Hack teaches you to find the single-vendor dependency you forgot you had. The medium is a runnable shell move; the domain is Workflow. Run it at the root of every repo your studio ships — it surfaces every place your code names exactly one LLM provider, which is exactly where Fugu’s spare path is missing.
git grep -nE "openai|anthropic|gemini|mistral|claude-|gpt-" -- '*.py' '*.ts' '*.js' \
| grep -vE "abstraction|router|provider" \
| sort -t: -k1 \
| awk -F: '{print $1}' | uniq -c | sort -rn
The files at the top of that count are your real dependency graph’s hot spots. Put a provider-agnostic layer in front of the worst offender before next week.
Looking ahead: Watch whether the orchestration idea (Fugu) collides with the open-file idea (CADAM) into a single demand — swappable models writing diffable, vendor-proof artefacts. Expect the secrets-hygiene and PRNG-provenance threads to harden into procurement language, and keep an eye on the wave-equation space-time solver: parallel-in-time FE is the kind of quiet infrastructure win that reshapes how Swiss acoustic and seismic teams budget compute this year.
PAZ Kaffi · multidisciplinary editorial, led by PAZ Academy