Redesigning the relationship between production and place.

From observation to action, at community speed.

Earth already has eyes on it — Copernicus, Earth Engine, Aurora, GraphCast. What it doesn’t have is the wire between those eyes and a network that can actually respond. PLANETAI is an instrument for closing that loop: community-scale sensing, bioregional decision, distributed fabrication response — in days-to-months, not years-to-decades.

Read the Core Ideas Paper Open the observatory

Trace the build-up to PLANETAI → 12 years, 16 projects, 4 phases

Duration: 36 months
Hypotheses: Two pre-registered
Bioregions: Four
Licence: Apache 2.0

The gap

Observation-rich.
Action-poor.

A satellite pass tells a European ministry that PM_2.5 in a valley crossed a threshold in March. The ministry feeds it into a policy cycle that reports in 2028, legislates in 2030, permits retrofits in 2032, and measures impact in 2035.

The observation was excellent. The response took a decade. During that decade the valley kept breathing the same air.

Existing Earth observation is calibrated to intergovernmental treaties. PLANETAI is built for the opposite loop.

PLANETAI’s nodes are also a way to protect distributed public capacity for observation, science, citizenship and policy when the centralised version is fragile.

Observation → action, two regimes

Policy cycle ~10 years

2025Observe

2028Report

2030Legislate

2032Permit

2035Measure

t = 0t + 10 years

Fab City cycle Days – months

d+0Detect

d+3Decide

d+14Fabricate

d+30Deploy

m+6Measure

t = 0t + 6 months

The instrument

A five-tier open AI stack,
bound to a standing network.

Each tier is simultaneously a computational layer, a governance layer, a data layer, and a fabrication-and-deployment layer. Dimensionality descends from planetary foundation models to on-device inference.

Fig. 1 · The descending stair (Full Stack Metrics scales) Dim / Tier

Planet 2048-d

Federated aggregation over open Earth foundation models — Aurora, Prithvi-WxC, GraphCast, GenCast.

Global summary

Bioregion 1024-d

Watershed / ecoregion. Partner-hosted servers, knowledge graphs, OSH/OKH response-template library, community-council veto.

Decision tier

Region 512-d

Multi-city / state / province. Where governance and supply-chain coordination actually live — Generalitat-class open-data publishers, regional GHG inventories, inter-city procurement.

Governance tier

City 256-d

Municipal dashboards, Metroverse ECI, Amsterdam Circular Monitor template, Boeing Fab City Index.

Urban accounting

Community 64-d

Neighborhood · fab-lab shed · campaign sensor archive. Compact federated-data node at the fab lab, local sensor integration, fabrication queue, action-cycle timestamps. PKC framework as the individual antecedent extended upward.

Fab Lab node

The node

PLANETAI is infrastructure, not an app.
The stack compiles to hardware. Four nodes ship.

The observatory you see is the cloud face of a deployable node — a small open-hardware box that lives at a fab lab and ingests planetary models, bioregional data, and its own community’s sensor feed. One node per pilot during the 36-month measurement window. On each node, a narrow set of action agents drafts measured responses into a queue — a named human approves or rejects before anything fires.

PLANETAI Node v0 · Open hardware

A federated-data node at the fab lab, not a datacentre.

The node runs the Region, City, and Community tiers as a storage-rich, asynchronous compute layer. It retains local data, caches what the bioregion needs, and pairs upstream Earth-foundation embeddings with locally-curated observations. Federated by design — one node per pilot, four nodes ship.

Compute: Compact general-purpose node · storage-rich · optimised for asynchronous federated workflows, not latency-sensitive edge inference · spec'd in the hardware annex
Sensors: Campaign sensor kit · environmental HAT · pilot-specific add-ons · deployed per campaign, not as continuous fleet
Licensing: Apache 2.0 / CERN-OHL · OSHWA-track certification

What the node does. Federated data retention, raster and sensor cache, model checkpoints, Region / City / Community-tier feature extraction, periodic model updates. Aurora and GraphCast run upstream at bioregion-server scale; the node consumes their embeddings and pairs them with locally-curated data.

Action agents · Human-in-the-loop

Four narrow agents, one per pilot.

Agents are constrained — each runs a single bioregion-specific workflow. They draft, they never dispatch. Every draft → approve → deploy cycle writes to a public audit ledger, timestamped against the action-latency hypothesis.

A1
Bali · Agent-1Wet-season + burning-trash air quality. Campaign-deployed kits + community reports + IQAir community feeds → drafts a ventilation / respirator-kit dispatch for Fab Lab Bali & CAST partners.
A2
Barcelona · Agent-2Circular-flow template scheduler. Fab Lab BCN waste-to-production loop → drafts weekly reuse / remanufacture tickets against the OSH/OKH library.
A3
Santiago · Agent-3Informal-economy visibility probe. Bridges the ~75% ECI blind spot — drafts survey sweeps and sensor-deploy tickets for Núcleo Milenio FAIR.
A4
Boston · Agent-4Community sensor QA + calibration. MIT CBA fleet → drafts calibration-run schedules and flags drift against co-located reference stations.

Not autonomous. Every action queued by an agent requires a named human approver at the fab lab. Governance is enforced at the bioregion-server boundary, not at the agent.

Two claims, on trial

Two pre-registered falsifiable hypotheses.
Positive or null, both will be published.

H₀−T · Throughput (null)

Does distributed production lower urban material throughput?

Null: distributed production at community and city tiers does not causally reduce urban material throughput or energy demand by more than 15% over 36 months, measured against a synthetic control baseline.

Three outcomes, all publishable. Strong positive (40–60%, LCAs suggest), partial positive (15–20%), or null. The in-window test is at community/district scale; bioregional H₀–T is a designed follow-up deliverable on month 60+ data when statistical lag and power both improve. A null at this scale would itself be a first.

H₀−A · Action-latency (null)

Does coupling observation to the network shorten response time?

Null: coupling multimodal observation to the Fab Lab network does not measurably shorten observation-to-action cycle time, relative to a policy-only baseline anchored on three completed Fab City Challenge editions (Bali 2022, Bhutan 2023, Mexico 2024; 12–16 weeks brief-to-prototype), across 4 community hubs (one per pilot) using difference-in-differences estimators.

Measurement. Five timestamped stages: detect → decide → fabricate → deploy → measure.

The network

The network is already here.
The wire is what’s missing.

~2,700

Fab Labs
Fablabs.io

Signatory
Fab Cities

15yr

Campaign archive,
Barcelona

Fab City Challenge
editions delivered

Apache 2.0
OpenRAIL
CC-BY 4.0

The instrument
is open

The leverage isn’t building a new network — the network exists. It’s the coupling that turns it into an instrument.

The monitoring case is in the literature: Balestrini, Diez et al. (EAI Trans. IoT, 2015) — a 15-month SmartCitizen deployment study across two communities — find local-champion orchestration is the determinant of sustained participatory sensing; Chen, González et al. (Atmospheric Pollution Research, 2025) show low-cost sensors deployed via fab-lab institutional support meet US EPA performance targets at R² > 0.7, RMSE < 5 µg/m³. The action case is also documented: Thomas (2023, ADB ex-ante BCA / Zenodo) projects $5–12 BCR for the 2023 Fab Bhutan Challenge through 2033.

Four bioregions

Calibrate once. Export honestly.
Falsify across climates.

Each bioregion discloses what exists and what’s being built. We do not flatten them into identical templates; instrumentation reality differs, and saying so matters more than a clean card grid.

Mediterranean P0 · anchor Pledge 2014

Barcelona

Instrumentation: strongest in the network. Live municipal feeds via Sentilo, 2,800+ open datasets via OpenData.cat MCP, plus a 15-year Smart Citizen campaign archive (~25 kits actively reporting on a given day — the campaign archive between campaigns). Full Eurostat EW-MFA, direct Metroverse coverage.

Where the stack is proven before it is exported.

Anchor partners IAAC · Fab Lab Barcelona · regional HPC & municipal partners under negotiation

North Atlantic P1 FAB26 + Summit host 2026

Boston

Cold-climate heat-island dynamics. Community sensor fleet needs building; we name that honestly.

First test of federation across a statistically different bioregion.

Anchor partners MIT Center for Bits and Atoms · Fab Foundation · Fab Hub Kendall Square

Southern Cone P1 Pledge 2017

Santiago

Andean / coastal / central-valley climatic bands. ~75% of urban employment is informal — ILO 21st ICLS (2023) is the Community-tier instrument.

Where formal-sector data alone produces a misleading picture.

Anchor partners Universidad Católica de Chile · Núcleo Milenio FAIR · Chilean AI research partners under negotiation

Indonesian Archipelago P2 Pledge 2022

Bali

Tropical-humid, cyclone-exposed, archipelagic. Bali falls into the same PITO/DIDO/ρ taxonomy as every other pilot. Two sovereign sources feed the framework: Bali Satu Data (4,700+ provincial datasets organised through Tri Hita Karana cosmological categories) for City- and Region-tier throughput, and the WIPO Global Innovation Index 2025 (Indonesia 55/139) as the country-level Bioregion peer anchor.

Where sovereignty-by-architecture is stress-tested hardest.

Anchor partners IT Del · CAST Foundation · Meaningful Design Group · Fab Lab Bali

What we’re honest about

Disclosures, upfront.

The observatory tags every value with provenance — LIVE, CACHED, SYNTHETIC, PENDING. The research programme holds itself to the same standard.

Connector pending
Cross-network sensor normalisation is not solved. It is a research deliverable of this proposal, not a completed primitive.
Synthetic proxy
Bali has no Metroverse City-tier ECI. We reconstruct the City-tier signal from Bali Satu Data (provincial open data, 4,700+ datasets, Tri Hita Karana taxonomy) feeding the same PITO/DIDO/ρ framework every other pilot uses — no Bali-specific index. Engineering ingests via data.go.id (CKAN 3.0 national portal) plus a Bali Satu Data scraper; partner collaboration with Diskominfos is year-one work.
Calibration required
Community sensors carry ±30–80% uncertainty against reference instruments (Castell et al. 2017); ≥30-day colocation is the required calibration protocol before any reading enters the bioregional index.
Commitment
We have committed to publishing null results if either hypothesis fails to reject. The pre-registration is the instrument’s own peer review.

Enter

Read the Core Ideas Paper.
Then open the instrument.

The Core Ideas Paper is the argument. The observatory is the first working prototype — partner-preview build, with live connectors already wired and synthetic placeholders clearly marked.