geo.qaAsk anything about the real world.
geo.qa is geographic question‑answering for production AI. A fleet of modalities (satellite optical, SAR, drone, CCTV, ground sensors) fused into one memory you own, so your model answers with what was actually observed, cites the receipt, and can train on it.
Chapter I
The agent that was never there.
Put an AI agent into production and, sooner or later, it answers a question about a place it has never observed. Fluently, and wrong.
If you have run an agent against the physical world, you have seen it. Ask whether an address flooded last week, whether a field caught rain, whether a gate stood open at three in the morning. The model reaches for the nearest static dataset and serves it back as live truth.
The failure is not reasoning. The model reasons fine. The failure is that it holds no memory of the place: nothing observed, nothing dated, nothing it can be held to.
This is a memory problem, not a model problem. So we built the memory: a record of real places, kept current by every sensor already pointed at them, that an agent can query and quote.
Chapter II
How the memory is made.
The encoder runs at the source, on our own devices. What travels to the ground is a latent, not imagery, and a decoder turns it into citeable fact.
Observe
Sensors already in place: orbital payloads, drones, vehicle cameras, fixed CCTV, all pointed at the same ground.
→Encode → latent
The encoder runs at the source: for satellites, on the orbital data-centre payload. A scene becomes a latent vector before it touches a ground station.
→Decode → memory
On the ground, the decoder resolves each latent onto the cell lattice: one shared, addressable Earth memory.
→File a fact
Every value is filed with its band, its time, and a signed receipt. Nothing is overwritten; the past stays queryable.
→Agent cites
Your model reads the memory and answers with the exact observations that support it, receipt attached.
Chapter III
One address for every place.
A memory is only useful if you can name a spot and trust the name. geo.qa gives every place on Earth a stable address and a simple algebra.
Chapter IV
Many modalities, one memory.
Every source writes into the same addressable record. The fleet is the unit of intelligence, not the single sensor.
| # | Modality | What it writes into the memory | Cadence |
|---|---|---|---|
| 01 | Satellite opticalmultispectral surface | land cover, water, change, vegetation indices | daily–weekly |
| 02 | Satellite SARall-weather radar | structure, moisture, flood, motion, through cloud | 1–6 days |
| 03 | Thermal / IRheat signature | flares, fires, equipment activity, anomalies | sub-daily |
| 04 | Drone surveycentimetre detail | close-range inspection of a single asset | on tasking |
| 05 | Vehicle camerasstreet level | what is round the corner, lane by lane | continuous |
| 06 | Fixed CCTVthe perimeter | gate, yard, fence line, occupancy | continuous |
| 07 | Ground sensorsSCADA / IoT | pressure, flow, level, vibration | seconds |
| 08 | Weather fieldsenvironment | rain, wind, temperature, soil moisture | hourly |
| 09 | AIS / vesselmaritime tracks | who is where on the water, and when | minutes |
| 10 | Field reportssigned human notes | a person’s observation, attributable | on event |
| 11 | Cadastral / GISthe record of record | parcels, assets, boundaries, ownership | on change |
| 12 | Acousticsound events | alarms, gunshots, machinery, leaks | continuous |
Chapter V
Where geo.qa sits.
Many are calling 2026 the year of the world model. Several layers of intelligence are converging, and each one needs something to be true about a real place.
| World & geospatial models | geo.qa | |
|---|---|---|
| Output | an inferred representation | an observed fact, dated |
| Trust | black-box inference | a signed, verifiable receipt |
| Deployment | a cloud service you call | your tenancy, airgapped |
| Time | a single snapshot | every tslot, append-only |
| Fusion | mostly one modality | many, one address |
| To build on | a fixed, pretrained model | a memory you train your own on |
World Labs’ Marble and DeepMind’s Genie generate a plausible world; they do not claim to reproduce the real one. Niantic’s Large Geospatial Model and Google’s AlphaEarth do read the real Earth, but as a model you call, not a memory you own, cite, and train on. geo.qa is the layer underneath: a dated, signed record of what was actually observed, private to you, that your own models can learn from.
Chapter VI
Train your own world model.
The memory is not only something to query. It is a substrate to learn from. Every dated, multimodal fact is supervision for a model you train, own, and deploy, without your data ever leaving the boundary.
# a JEPA dynamics model, trained on your own memory from geoqa import Memory, train mem = Memory(tenant="acme") model = train.jepa( memory = mem, bands = ["optical", "SAR-VV", "thermal", "weather"], horizon = "7d", ) model.deploy(airgapped=True) # reads the same memory
Chapter VII
Airgapped by construction.
Because what moves is latents and signed facts, and the memory is a tenancy you own, your model can run sealed, train sealed, and still be audited.
What never leaves
the boundary holds by design, not by policy
- ◇Raw imagery. Encoded at the source; only the latent vector travels down.
- ◇Your memory. A single tenant: your keys, your retention, your jurisdiction.
- ◇Your models. Trained and served inside your boundary. No calls home, no training on your data by us.
What you keep
provenance that survives a hostile review
- ◇Provenance. Every answer carries a signed, verifiable receipt.
- ◇Audit. Replay any decision against the exact observations it used.
- ◇Interop. The same call shape as the open emem.dev protocol.
Chapter VIII
An interface agents already speak.
One call shape, shared with the open emem.dev protocol, so an agent reads public and private memory in a single answer.
# resolve a place, ask a question, get a cited fact from geoqa import Memory mem = Memory(tenant="acme") # your private tenancy cell = mem.locate(lat=25.276, lon=55.296) fact = mem.ask( cell, band = "flood_extent", since = "-30d", ) fact.value # 0.0 fact.sources # 4 observations · 2 modalities fact.receipt # 0x9f3c… signed · verifiable
Chapter IX
Questions from the field.
Most operators feed the encoder and read the decoder at the same time. A few of the surfaces we keep ending up on.
Autonomous fleets
The vehicle’s cameras write to the memory; the car reads neighbouring cells to know what is round the corner before it sees it.
what is in cell 8a2a1072b… right now?
Energy & infrastructure
Pads, lines and stations watched by every sensor at once. One page goes out only when independent witnesses agree.
did pressure at G-247 drop overnight?
Insurance & risk
Price and settle against what was observed, not what was claimed, and every fact arrives with its receipt.
was this parcel underwater on 04 May?
Defence & GEOINT
An owned, airgapped memory with full provenance: answers built to survive an audit, weeks later.
what changed at this site since Tuesday?
Give your model a memory.
Stop letting it guess about the world. Start letting it cite, and train on what it cites.
geo.qa · a vortx ground decoder · emem.dev open protocol