Three horizons. Each workstream has a definition of done; nothing here is aspiration without a shipping shape. Dates are calendar quarters and mark intent under current resourcing — design-partner work always outranks the roadmap when they conflict. This is the product plan; the company-level timeline — revenue targets, hiring, funding, and the gates where course changes — is the business plan.
We plan in independently shippable slices and have held that cadence all year:
| Shipped | What |
|---|---|
| 2026 Q2 | World-action layer (stochastic, skill-gated outcomes) closing the calibration loop; multi-infrastructure model connectors; operator dashboard with live map and ops firehose |
| 2026-07-07 | “The Ghost in the Shell” — public talk and paper argument (repo) |
| 2026-07-09 | Ghost Gateway owner keys · ghostkit · GhostDeck desktop builds for three platforms |
| 2026-07-10 | geobard extracted as a standalone OSS project |
| 2026-08 | ticker (label-driven world-time service) · avatars (in-character chat API, live for the OHM ruler galleries — the engine’s first non-game deployment) |
The commercial gate for everything else: convert studio conversations into evaluations, and evaluations into design partnerships.
| Workstream | Deliverable | Done when |
|---|---|---|
| Evaluator sandbox | A curated docker compose up preset: one command brings up engine + organs + dashboard + a seeded world, with a documented quickstart and reference-architecture diagrams for Unity, Unreal and Foundry pipelines |
A studio engineer reaches a living, inspectable world in under 15 minutes without talking to us |
| Public proof surfaces | Hosted Object Inspector demo (description in → structured affordances out, latency visible) and an influence-ripple demo on a staged scenario | Both linkable from the site and stable enough to leave running |
| World richness | Failure-aware world verbs (climb, pick-lock, persuade, trade) resolving through the same stochastic layer, so calibration outcomes become richer than success/failure of dispatch | Calibration records carry world-grounded outcome detail |
| Population scale | From tens to thousands of concurrent agents: template-driven mass spawning, affordance-aware scheduling of agent workloads across hosts, and the load engineering that follows | A thousand-agent world runs for a week unattended with a flat operator cost curve |
| Self-serve ownership | Owner-key issuance without an operator in the loop, abuse controls (creation flooding; energy already throttles action spam), and a custodial-runner convenience tier | A stranger can own and run a ghost without us provisioning anything |
| Partner integrations | 2–3 design partnerships; world-data ingestion adapters for partner-owned worlds | Partner world state served by the engine in their pipeline, LOI-backed |
Horizon 1 sells to CTOs; Horizon 2 is what their creative directors use. This is where the paid layer becomes visibly more than the OSS sum-of-parts.
The calibration programme is the scientific spine: extend the affordance-order overconfidence results across model families and quantisation levels, test mitigation strategies (two-tier vetting already measurably reduces failure dispatch), and publish against the controlled benchmarks the paper targets. Research output is marketing, hiring surface, and product regression suite at once.
| Risk | Posture |
|---|---|
| Studio evaluation cycles are slower than our runway math | Horizon 1’s sandbox exists to compress them; the OSS believer-funnel keeps inbound warm between BD pushes |
| A model provider ships a “world memory” feature that sounds like us | Their layer is per-conversation state; ours is authoritative, multi-agent, designer-governed world law. The category work in 03 is the defence, and it’s why we refuse the wrapper shelf |
| A clever team rebuilds the stack from our published architecture | Some will; they’re lost-and-fine. The maintained engine, the world adapters, the designer tools, and the compounding calibration data are the multi-year bet they’d be making against |