Products & Publishers
Expands Deliverable 6 §2–3 of the research pass. The buyer-value logic per quest lives in Quests as a Data Product; this page is about what we sell, to whom, in what order, and why publishers move from risk column to asset column.
Design principle
Sell the engine's outputs in the order of least legal exposure and least corpus-size dependence first. Raw footage licensing is deliberately last, and only ever on publisher-partnered titles.
The open-platform reformulation (2026-07-08)
The "Open Data & the Spatial Wiki" vision (Guido, post-Sestini call —
research/world-model-data-flywheel/vision-open-data-spatial-wiki-2026-07-07.md)
reframes these products as the paid tier of an open platform: researchers
browse open datasets and open requests ("a game, a place, a behavior");
players fulfill them by playing; the same sessions build a community-owned
spatial wiki. Product 1 is that request loop productized — Sestini and
David independently invented it as their dream tool, making five independent
convergences on directed collection. Openness is the growth mechanism (open
datasets → papers → replication demand) and dissolves the Hugo
synthetic-objection, provenance anxiety, and AI-Act disclosure at once; the
spatial wiki is the community counterweight that keeps directed collection
from becoming "work."
Team decision #7 — the open/owned split (strawman in
ec-workspace/research/world-model-data-flywheel/vision-integration-memo-2026-07-08.md):
OPEN = protocol + probe suite, a Tier-1 commodity slice per title, FS-001
format + loaders, the naturalness measurement. OWNED/PAID = the eval
instrument as a service, gold-tier curated slices, directed campaigns as
priority (paid requests jump the queue; open requests ride community
capacity), longitudinal histories, the contributor network. Open sub-question
for counsel: whether time-boxed slice exclusivity survives at all, and which
license open slices carry per title tier. Launch scope is unchanged — the
vision is the horizon, the quest+benchmark path is the road.
Product 1 — Directed-collection-as-a-service
What it is: a buyer (world-model lab, later a studio) specifies eval gaps — weak scenarios, thin slices, missing distributions. Our quest compiler prices them, the fleet fills them, and delivery includes the measured coverage delta, not just files.
Why first:
- It monetizes the flywheel, not the archive — so it works while the corpus is small. Corpus size is optimized last (Roadmap).
- It is the one thing the market leader structurally cannot offer (Medal: no quests, no raw inputs) and the sub-scale rival can't do credibly yet (Grunt).
- It naturally produces early-access-style contracts — the revenue template General Intuition proved (~$40M reported 2025, thinly sourced but directionally confirmed by their disclosed "handful of customers" model).
- Legally cleanest posture short of publisher partnership: bespoke collection under explicit contributor consent, scoped to whitelisted titles, delivered with provenance.
Pricing: per verified coverage unit against the working band — now anchored by EC's own real numbers (Roblox: $8–10/hr verbal for commodity volume; $70–85/hr effective on the structured first batch via NRE separation; €65/kept-hour is the internal economics anchor) alongside the research comparables ($60–240/hr footage → $118–200/hr teleop). The measured delta justifies the top of the band; the founding experiment supplies the first delta number. Compliance note: spec'd task work under electronic supervision is exactly the EU Platform Workers Directive's algorithmic-management pattern (Risk & Compliance #10, transposition 2026-12-02) — employment counsel reviews the campaign structure before the first paid engagement.
Product 2 — The action-inference model + labeled/curated slices
What it is: an inverse-dynamics model trained across the whitelisted catalogue (the VPT/D2E pattern), licensed as an action-inference model the buyer applies to their own video problem (external naming note: "labeler/annotation" is banned register in EC external copy); plus curated, provenance-clean data slices selected by the coverage atlas.
Positioning constraint (vault, 2026-07-08): IDM-inferred actions are the cheap substitute EC's ground-truth premium is defined against — so this product only coheres as "we own both tiers and publish the measured delta": the cross-game adaptation curve is simultaneously Product 2's pricing model and the ground-truth premium's evidence. Never market it as equivalent to captured ground truth.
Why second:
- Highest leverage per collected hour: ~2K labeled hours historically unlocks 70K scraped hours (35:1). The IDM is the concentrated form of the corpus's value.
- Cleaner IP posture than footage resale: the buyer receives a model (or annotations), not copies of publishers' audiovisual works at scale. This matters — see Risk & Compliance #1.
- Differentiated where reconstruction fails: KBM-precise titles (high-DPI aim) and voice-aligned intent, which NitroGen-style overlay parsing cannot touch.
- It compounds with Product 1: every directed campaign improves the IDM; the IDM cheapens every future campaign's labeling cost.
The research milestone that gates it: IDM labels held-out scraped video of an unseen game above a usefulness bar (VPT's 90.6% keypress / 0.97 mouse R² is the in-distribution reference; cross-game transfer is the research question, and beating D2E's Generalist-IDM on KBM titles is the concrete target).
Product 3 — Playtesting agents & coverage analytics for studios
What it is: internal use of our own data — agents that explore builds, find crashes/exploits/coverage holes, and report; analytics on how real players traverse a title, where they get stuck, what they never find.
Why it matters strategically (beyond revenue):
- It's the nearest budget line: studios already pay for playtesting (PlaytestCloud's 1.5M-player panel proves the market), and it doesn't wait on any AI-lab procurement cycle.
- It is the publisher flywheel's entry point — see below.
- It exercises the whole engine on a friendly customer: eval slices, quest targeting ("get 50 players through the new zone tonight"), delta reporting.
The publisher flywheel
The problem it solves: contributor consent never clears the publisher's copyright in the game's audiovisual output — the #1 risk and most plausible deal-breaker. The strategic answer is not to minimize contact with publishers but to make them customers, then partners.
The end-state dataset nobody has: server-side ground truth (from the partner studio's engine/backend) paired with client-side inputs and POV (from our fleet), cross-publisher. Microsoft has it for first-party only; Origin Lab has engine-side without a player fleet; we'd have both. This is the 3-year moat candidate #3 from Competition & Moat.
Sequencing note: the flywheel starts with one studio (indie/AA, PC-first, no kernel anti-cheat, EULA silent-or-friendly on recording) — not with EA. The Blizzard StarCraft II AI/ML license precedent shows publishers already treat ML use as separately licensable; our pitch makes granting it profitable.
What we do NOT sell
- Raw footage of non-partnered titles. The single highest-exposure act (Bartz library-side; Valve's "never sell your videos" clause; Frontier-style AI bans). Whitelisted-title footage may be bundled within Product 1 deliveries under buyer-side risk allocation — but footage-as-catalogue waits for publisher partnership.
- Manipulation/force "robotics" data. Fake-transfer (Technical Evidence §C; Quests as a Data Product verdict rule). Robotics buyers get navigation/locomotion/coordination slices, honestly labeled as high-level priors, when they show up with budget — not before.
- Anything on non-consented or under-18 capture. Non-negotiable (Risk & Compliance #3).
Revenue shape over time
| Horizon | Dominant revenue | Corpus role |
|---|---|---|
| Phase 1–2 (0–4 quarters) | Product 1 early-access contracts (1–2 labs); first studio engagement | Small, whitelisted, experiment-grade |
| Phase 2–3 (3–8 quarters) | Product 1 recurring campaigns + Product 2 IDM/slice licensing; studio analytics recurring | Growing against demand, never ahead of it |
| Phase 3+ | Publisher-partnered premium datasets (both layers cleared); Product 3 at portfolio scale; Leg B option if it matures | The rights-cleared subset is the premium tier; longitudinal histories compound |