Archive · June 2026
Voronoi: the pattern a body uses to spend the least material
A PAZ foundation essay on Voronoi diagrams — the geometry of nearest-neighbour packing, its Delaunay dual, and how it zones CLT, lattice infill and façades.
Make Your Own Website: The Plain-HTML Stack That Still Opens in 2070
A hands-on PAZ tutorial: build a website from one hand-written index.html with VSCodium — the durable, vendor-free stack that still opens decades from now.
The self-certifying cache: why LAWS could make on-site robot AI provable
LAWS proposes an inference cache with a deployment-time error bound you can check without ground truth — what it means for on-site robots and BIM AI tools.
The Building's On-Premise Brain Hallucinates Its Own API
PowerCodeBench shows on-premise LLMs hallucinate pandapower APIs — demand-guided docs lift accuracy 32–56 points at 41% token cost. Building-mind angle.
Talos on Hetzner: A Private Kubernetes Your Successor Can Still Boot
Build a private, EU-hosted Talos Kubernetes cluster on Hetzner with Terraform — no public IPs, reproducible from git. A hands-on PAZ Academy tutorial.
One Hyperbolic Element Decides If Your System Holds Steady
A May 2026 arXiv paper reduces a hard dynamics question to one algebraic test — is there a hyperbolic element? — and shows why the rigid body can't comply.
A Novelist Replaced Adobe with Git. Architects, Draw Your Dependency Graph.
DJ Speckhals replaced Word, InDesign, and Kindle Create with Pandoc, Typst, and Git. Why his dependency graph is a warning for every Autodesk shop.
CAD: the program behind every building, and the question of who is still drawing
A PAZ foundation essay on CAD — from Sutherland's 1963 Sketchpad to BIM and code-CAD — and why the model, not the drawing, is the real deliverable.
Review of the Week: The Building Reads Its Own Dependency Graph (2026-W23)
Week 23, 2026 recap: PINN robot control, VTT's grabbing truss, MAGS-SLAM, and the one move that defined the week — draw your real dependency graph.
Basel's Tail Metric Taught a Neural Net — and Your Studio Has the Same Scarce-Data Problem
A new arXiv paper distils a CVaR risk optimizer into neural students from 104 samples — and the teacher-student trick maps straight onto AEC's data gap.