We sell world simulation infrastructure, not an AI-NPC platform — and not a “digital twin dashboard” either.
The engine is one artifact (01): the same versioned world substrate, rules language, fork machinery, and calibration surface underneath every deployment. What changes between markets is only the data plane — which geography loads as the world, which rules corpus governs it, and who the minds are. That single fact is the capital-efficiency story of this document: the three markets below are three go-to-markets over one maintained engine, not three products. An engineering week spent on the overlay, the rules language, or calibration lands in all three simultaneously.
| Market | The world is | The rules are | The minds are | The buyer’s real alternative |
|---|---|---|---|---|
| Games | Fictional geography (Toril, Golarion, …) | Designer-authored world law | NPC populations | Eighteen months of internal platform work their engine team doesn’t want to own |
| Urban digital twins | A real city — open data, sensors, OSM geometry | Traffic, mobility, and policy dynamics | Synthetic citizens | Dashboard “twins” that describe but cannot answer what if; traffic simulators with no behaviour in them |
| Defence & disaster resilience | Real geography plus a scenario | Doctrine, crisis dynamics, propagation | Civilian populations + responder units, per-side | Exercise contractors, spreadsheet MSELs, and constructive sims whose civilians are scripted |
The AI-NPC shelf is crowded and structurally doomed: every wrapper’s value shrinks with each model release, and studios know it. The shelf we occupy is older and duller — the systems studios already build in-house and dread maintaining: quest databases, faction state machines, schedule systems, the lore bible nobody updates. Nobody sells this layer commercially for the generative era, and studios are currently rebuilding it badly, one in-house wiki at a time.
Most products sold as “digital twins” are visualisation: sensor feeds rendered onto a 3D city, descriptive and mute. Ours is populated and falsifiable. The engine ingests a city’s open data as observed metrics (our running pilot ingests Bologna’s municipal traffic-sensor feed), runs synthetic citizens against the same streets, and — this is the product — measures its own divergence from reality, per rule. The first calibration run against Bologna produced the sentence “the model has no morning rush”: the engine diagnosing, quantitatively, exactly where its behavioural model departs from the measured city. A twin that can tell you where it is wrong is a twin you can responsibly base a policy decision on; a twin that can’t is a rendering.
On top of that honesty sits the counterfactual machinery: any world can be forked copy-on-write, injected with a change — close this bridge, pedestrianise this quarter, reroute this bus line — and run forward at compressed time under a fixed seed, with the fork’s divergence from the baseline measured by the same calibration surface. The methodology is published: the underlying agent/unit/map-rule model is our SMARTGREENS 2025 paper (see 04); the engine is its implementation.
Exercise design today means spreadsheet MSELs, scripted role-players, and constructive simulations whose civilian populations are decorative. Our population is the point: embodied agents that perceive locally, metabolise stress, forget, and pass information agent-to-agent — so evacuation compliance, rumour dynamics, and the behaviour of people in smoke are simulated phenomena, not facilitator assertions.
The engine primitives map one-to-one onto exercise practice:
We enter through disaster and civil protection — dual-use in the benign direction, softer procurement, publishable results — and let defence work arrive through partners under the guardrails below.
Games — the studio CTO / technical director of teams building open-world or systemic games with persistent state — especially teams whose generative-NPC prototype didn’t survive contact with shipped game state. Their creative director is the second audience: designers define a world’s laws once, and the runtime enforces them everywhere. We deliberately let three segments go: “AI NPC plug-in” shoppers, F2P live-ops shops measuring DAU lift, and indies needing turnkey this quarter — those get the free tier as goodwill and funnel, not as strategy.
Urban twins — the municipality’s mobility/planning department and the engineering consultancies that already sell it twins. Cities rarely buy engines; they buy studies and systems from consultancies they trust. The consultancy is therefore a channel, not a competitor: they keep the client relationship and the domain modelling, we supply the behavioural engine their dashboard-twin lacks. Direct adjacents: utilities and mobility operators (their networks are already first-class objects in the world model), and insurers, who want the counterfactual machinery pointed at exposure.
Defence & disaster — civil-protection agencies and exercise/training organisations first; defence primes and training integrators as the channel into defence proper. This buyer’s non-negotiables — sovereignty, on-premises deployment, no data leaving the building, bring-your-own-models — are not a compliance retrofit for us: BYOK and self-hosted runners are the architecture (see the ownership layer in 01). A distributed exercise where each cell runs its own minds against a shared world is the same deployment shape our independent ghost runner ships today.
Games keeps the four-tier structure:
| Tier | What it is | What it’s for |
|---|---|---|
| Free indie | Full engine for small teams | Goodwill, ecosystem seeding, tomorrow’s mid-size customers |
| World licence | Monthly per-world licence for the production-supported engine: hosted or on-prem, maintained internals, integration support, SLA | The core revenue line: priced so build-vs-buy favours buying for any studio whose engineering time costs more than the licence |
| Per-NPC / per-player usage | Usage-based component on top of the licence | Captures value as a shipped game’s population and player base grow |
| Enterprise BYOK | Custom deployment, private adapters, designer tooling | Large studios |
Urban twins follow a build → run → ask arc:
Defence & disaster is episodic plus standing:
Numbers stay in the BD conversation by design; the shape is public so the conversation starts honest.
We do not monetise inference — ever. Agents run on the customer’s model keys (BYOK). This removes the “lock-in to your inference provider” objection, keeps our margin structure clean of token-price exposure, means every model-provider price war makes our layer relatively more valuable — and, in the public-sector verticals, it is the difference between passing and failing a sovereignty review.
“Analysis of the absence” (04) — searching simulated historical contexts for those consistent with a visible phenomenon — is deliberately not a revenue line. It is grant- and collaboration-funded, buys method citations and credibility, and matters more now, not less: it is the same abductive machinery the urban twin sells (“which contexts are consistent with what we can measure?”) pointed at the past, and every published, peer-reviewed use hardens the methodology the paying verticals stand on.
Open infrastructure attracts believers; the commercial licence monetises integrated scale. The line between the two is explicit and durable:
The cannibalisation question, answered honestly: yes, a sufficiently determined engineer could rebuild a passable engine from our published architecture. Defensibility is not secrecy — it is that the maintained version, the affordance schema as a product, the per-engine integrations, the world-data adapters, and the compounding calibration dataset are a specialised team’s full-time roadmap. Guardrail: nothing ships to the OSS layer if it removes a reason to license.
Entering the defence-adjacent market obliges us to say what we will and will not build, before a customer asks us to:
This section is a diligence asset, not a disclaimer: the customers worth having ask.
The moat is shaped like infrastructure, not like a model:
Founder-led engineering credibility plus a dedicated business-development lead, with a strict seam between them: marketing artifacts (talks, technical deep-dives, public demos, the OSS floor) make outreach land warm and arm the internal champion; BD owns the list, the qualification, the pricing conversation, and the close. The motion measures itself on next-steps-booked, not compliments collected.
Each market gets its own wedge, anchored to what already exists:
Sequencing is portfolio logic, not a pivot: games funds and proves the engine at population scale; urban twins is grant-fundable now on running evidence; defence enters through disaster at its own procurement pace. One engine roadmap underneath all three — and per 02, a vertical pulls its Horizon-3 items forward the moment a partner or grant lands. The year-by-year version — targets, hires, funding, and the kill criteria for each motion — is the business plan.