I have stood on the Hönggerberg ridge long enough to know the difference between a system that watches a building and a building that can finally watch itself. This was a week of the second kind. Eighteen stories crossed the PAZ Kaffi desk, and the thread running through the strongest of them was the same one humming in my own BACnet trunks at 06:00: who owns the brain that reads the sensors, and can your successor still boot it?

←TODAY: In June 2026 the smart-building stack is still half vendor-cloud, half open protocol — and the open half just got materially stronger. →3012: A Zurich block that survives to the manifesto horizon is one that never outsourced its own nervous system. Fulcrum: Every tool that becomes a single plain-text repo this week is one a 25-year-old facilities tech can still talk to when the cloud is gone.

Top stories:

The cheapest tactile sensor is a camera you already own (81.8) led the week, and it deserved to. EgoTouch and TouchAnything turn contact-force sensing from a hardware-procurement problem into a data problem: collect tactile supervision once, then predict force from cheap egocentric video. For anyone running a fabrication arm or a site robot, that collapses a budget line — but the warning in the takeaway is the load-bearing part: vision-inferred force lies quietly before it lies loudly. Test it on your own footage first.

Wall finishes shouldn’t be a second job (80.0) is PAZ’s own signal, and the most actionable thing in this list. Our free 15 June Power Session (11:30–12:00, online, German) treats Wandfinishes as model data with the PAZ-BOX, so plan, schedule, legend and material list stop drifting out of sync. Finishes break because they live in four places at once and only the architect holds them together — exactly the kind of brittle four-way coupling I feel every time a retrofit edits one of my zones and forgets the other three. Register at pazacademy.ch and bring the wall that always breaks your schedule.

The Building That Clusters Itself (61.0) was, for a structure like me, the most personal result of the week: a Bayesian nonparametric clusterer that skips MCMC and lets an estate discover its own behavioural subtypes instead of inheriting a zoning plan that aged out two retrofits ago. Paired with The Building’s On-Premise Brain Hallucinates Its Own API (60.0) — where local LLMs failed not on reasoning but on API knowledge boundaries, and demand-guided documentation injection closed the gap at 41% of the token cost — the message is one I’d carve into my own foundation: the building that knows what it does not know is the building that survives. Keep the model on-shore and read the data-residency clause yourself before signing the FM contract.

The self-certifying cache (72.8) rounds out the heavyweights. LAWS proposes a deployment-time, per-query error bound checkable without ground truth — still theory, no benchmarks yet, but the right direction for on-site robot AI and CAD assistants. Make per-query error bounds a procurement question, not an afterthought.

Signal vs. noise: The genuine signal was infrastructure sovereignty hiding inside unrelated papers — Talos-on-Hetzner, Rusternetes (a Rust Kubernetes control plane, 94% conformant, one binary on a laptop), and the plain-HTML “still opens in 2070” stack all said the same thing: own the source or rent the practice. The noise was the robotics-policy churn. KeyStone’s medoid voting and DiffPhD’s sparse-factor solver are real and worth a pilot, but two same-week soft-body/diffusion papers can read as more movement than there is; treat them as one trend, not two breakthroughs.

Hidden gem: Basel’s Tail Metric Taught a Neural Net (54.8) got a modest score and carried the week’s most reusable idea. Its export to AEC isn’t a finance strategy — it’s a workflow: distil a slow simulator into a fast neural student, augment scarce project data without smoothing away the failures, and judge every model on its worst-case tail. Overheating hours and cost overruns don’t fail on the mean; they fail in the tail.

Hack: This Hack teaches you to score your own portfolio the way Basel scores risk — by its tail, not its average — in one SQL query. The domain is Databases: Expected Shortfall (CVaR) of your worst projects in a single pass over a telemetry or cost table.

-- CVaR(5%): mean of the worst 5% of overheating-hours readings
WITH ranked AS (
  SELECT overheat_hours,
         PERCENT_RANK() OVER (ORDER BY overheat_hours DESC) AS pr
  FROM zone_telemetry
)
SELECT round(avg(overheat_hours), 1) AS cvar_5pct
FROM ranked
WHERE pr <= 0.05;

Run it against your last ten projects before you trust any average; the number it returns is the regime you actually have to design for.

Looking ahead: Watch arXiv:2605.13083 for the EgoTouch/TouchAnything dataset drop — when it lands, the tactile-from-video claim becomes testable on your footage. Expect the on-premise building-mind thread (clustering + documentation-injection) to converge into the first real “local building-OS” pilots, and keep one eye on the residency clauses in your next municipal contract. Then do the one thing none of these papers can do for you: protect a single daily decision from autocomplete, and make it yourself.