Give artificial minds real places to live — and keep the minds private property of their owners while the places stay shared.
Two convictions drive everything we build:
The second conviction is our worldview as much as our architecture: open infrastructure beneath commercial scale, non-extractive by construction. We monetise the maintained engine at the integration layer, never the inference, never the owner’s data, never attention.
GaiaWM was not started as an AI company. It grew out of a decade of open geospatial work:
That lineage is why the stack is geospatial-native (most agent frameworks bolt “location” on as a string), why temporal state is a first-class citizen, and why our first non-game deployment — in-character historical avatars over Open History Map’s ruler galleries — took days, not months.
The architecture is also an experimental instrument. Our thesis work — The Ghost in the Shell (presentation, 2026-07-07) — argues that embodiment constraints (position blindness, energy metabolism, decaying memory) are not game flavour but the substrate cognition requires, and that agent quality must be measured, not vibed: every prediction an agent makes is scored against the real outcome in the calibration organ.
The headline empirical result — overconfidence compounds with affordance order, reproduced live in our own simulation (+0.04 → +0.18 → +0.32 bias across orders 1–3 over 600 trials) — is exactly the failure mode that makes naive LLM-NPC deployments collapse in shipped games, and the two-tier vetted architecture is our measured mitigation. Research, product regression suite, and marketing are the same artifact here.
A persistent Toril (Forgotten Realms) simulation runs continuously on the full stack: LLM-driven populations spawned from declarative templates (farmers, bandits, travellers), named characters with souls and goals, ghosts running on everything from 1-bit-quantised local models to frontier APIs across six inference providers, an operator dashboard streaming every organ call and LLM exchange live onto a real map, and owner-run minds connecting through ghostkit and GhostDeck from outside our infrastructure. It has produced over 10,000 calibration records. It is not a demo reel; it is where we live.
Games are the proving ground, not the perimeter. The engine’s primitives — real geography, temporal state, embodied populations, propagated consequence, measured behaviour — carry directly into three directions we are actively developing. The first two are now markets in their own right in the business model; the third is deliberately not.
The methodology predates the engine: our SMARTGREENS 2025 paper (Montanari & Costagliola) defines a GlassBox-style simulation model — units carrying metrics, agents moving between them, map-level dynamics, rules written in a small declarative language — for city-scale digital twins. The engine is that paper implemented: the same rules language now drives the live simulation, and the paper’s own listings run as its conformance tests.
The pilot is running: the engine ingests Bologna’s municipal traffic-sensor open data as observed ground truth, simulates synthetic commuters on the same street geometry, and scores the simulation against the sensors per rule. The first calibration run reported that the behavioural model “has no morning rush” — a measured, quantitative statement of exactly where the twin diverges from the city. That sentence is the product: a twin that can locate its own wrongness can carry a policy question; a twin that cannot is a rendering. Counterfactuals come from the same machinery that runs game worlds — fork the city copy-on-write, inject the intervention, run forward seeded and time-compressed, measure the divergence.
Civil-protection scenarios — floods, evacuations, infrastructure failure, mass events — and exercise-driven defence training, simulated on the real world’s geometry with populations of embodied agents. The stack is unusually honest about how people actually behave in a crisis, because its constraints are the point rather than patches:
The engine’s exercise machinery maps onto practice directly: per-principal views give each side or agency its own honestly-lagging picture of the shared ground truth (fog of war, or a Common Operating Picture); injects are changesets under the exercise’s own authority; forks are courses of action, run seeded and time-compressed for measured comparison; and because the world is bitemporal, the after-action review — who believed what, when, on whose authority — is a query, not an archaeology project. Commercial shape and the dual-use guardrails we hold ourselves to are in 03.
Historical research is full of situations where the phenomenon is visible but the context is not completely explored: the record shows an outcome — a settlement abandoned, a trade route gone quiet, a practice that spread or failed to — while the everyday context that produced it left few or no sources.
The engine turns that absence into a testable target. Rebuild the period world from Open History Map data (geometry plus timeline), populate it with embodied agents under period-plausible constraints — skills, affordances, energy costs, travel and communication ranges — and search the space of candidate contexts for those whose simulated dynamics actually reproduce the visible phenomenon. Simulation as an abductive instrument: not “what happened”, but “which contexts are even consistent with what we can see”.
The calibration organ is what separates this from storytelling — every run is measured, comparable, and repeatable, which is the methodological bar academic use demands. We run this as a deliberately academic branch of the platform: university collaborations, grant-funded projects and publications, valued in credibility, method citations and calibration data rather than licence revenue.
| Material | Link |
|---|---|
| Pitch deck (live) | gaiawm.github.io/pitch-deck · source |
| The Ghost in the Shell — talk | github.com/GaiaWM/260707 |
| Organisation page | gaiawm.github.io |
| ghostkit — own a ghost | github.com/GaiaWM/ghostkit |
| GhostDeck — desktop client | github.com/GaiaWM/ghostdeck |
| Open-infrastructure floor | github.com/openfantasymap |
| geomqtt live demo | openfantasymap.github.io/geomqtt |
| Live tile infrastructure | fantasymaps.org |
For anything held back from this room — internals, numbers, partner material — open an issue on this repository or reach us through the organisation page, and we’ll take it to the NDA conversation where it belongs.