Passa al contenuto principale

Overview: The World-Model Data Flywheel

Status — strategy series, grounded in research

This series turns the July 2026 deep-research pass (ec-workspace/research/world-model-data-flywheel/, six parallel research agents + adversarial verification of 22 load-bearing claims against primary sources) into the strategy layer above the quests thread. It answers why the data engine is the business, where Quests as a Data Product answers which quests to build, Quests — Technical Design answers how they're built, and the Reward Economy answers how they pay. Nothing here is shipped; everything here is falsifiable and says so.

Revised 2026-07-08 (alignment reviews vs prior docs, analytics, and the EC Vault — see alignment-review-2026-07-08.md + vault addendum): buyer priority corrected toward frontier-lab/publisher procurement and EC's named live threads; capture spec corrected to built reality; coverage system merged with the prior realtime-spatial-coverage work; probes extended (P9/P10); fleet-scale assumptions grounded. Language rules apply throughout: EC says "action-conditioned world models" (never "world simulator"), never "annotation/labeling" externally, no "winner-takes-all" framing, and capture-title names never appear on public surfaces (activity-not-title rule) — this series is internal.

The hypothesis under investigation

Color legend — evidence grade per arrow (details in Technical Evidence):

ColorMeaningExamples
🟢 GreenStrong evidenceCapture at $20/hr works (VPT); every playable neural engine needed action labels; the market got priced (Medal/$500M offer, General Intuition/$2.3B)
🟡 AmberSuggestiveDirected collection beats passive volume where measured (WHAMM, Waymo +31%, VPT 35:1 leverage) — never yet for cross-game world models
🟠 OrangeA betGames→robotics transfer (one thin datapoint); the full human-fleet loop (nobody has closed it)

The strategic position

Playroll is not a data brokerage and not a frontier lab. It is a data engine — proven end-to-end on small models — selling the three things that fall out of it.

The three products, in order of shipability (expanded in Products & Publishers):

  1. Directed-collection-as-a-service — a buyer specifies eval gaps; the fleet fills them; delivery includes the measured coverage delta. Monetizes the flywheel, not the archive, so it works while the corpus is small.
  2. The labeler (Generalist-IDM) + labeled/curated slices — the highest-leverage artifact the corpus produces, and a cleaner IP posture than reselling footage.
  3. Playtesting agents & coverage analytics for studios — internal usage that produces revenue and converts the #1 legal risk (publisher IP) into partnerships.

Raw footage licensing is deliberately last, and only on publisher-partnered titles.

Why not a brokerage

Pure data resale is the weakest position in this market: a handful of buyers, no established price, and the publisher-IP risk falls hardest on exactly the act of reselling footage (Risk & Compliance #1). The one company holding the best corpus — Medal — chose to become a lab (General Intuition) rather than sell. This is the same conclusion Quests as a Data Product §0 reached from the product side: we are not selling gameplay video; we sell dense frame-synced actions, verbalized intent, multi-sync POV, and clean provenance.

Why not a frontier lab

General Intuition has raised $454M; Decart sits at ~$4B; DeepMind ships Genie 3. Competing on frontier world models is capital suicide. But nothing about proving data value requires a frontier model: Microsoft's WHAMM was rebuilt real-time on one week of curated data; VPT's inverse-dynamics model was small; open models (Matrix-Game, MineWorld) can be fine-tuned to demonstrate deltas cheaply.

Why the data engine

Nobody has published "coverage-directed cross-game collection beats passive volume for world models." Microsoft (WHAMM), Waymo (+31% on rare slices from 3% targeted additions), and OpenAI (VPT's 35:1 labeled:unlabeled leverage) each proved a piece of it in adjacent domains. The full loop — eval → quests → human fleet → measured model delta — is unclaimed territory: Tesla closes it with on-device triggers (no humans directed), DeepMind's SIMA 2 closes it with AI agents in generated worlds (no human fleet), and the data vendors (Scale/Surge/Mercor) close it manually. Playroll generating the first direct evidence is simultaneously the pricing mechanism, the pitch deck, and the publishable moat.

The two legs, honestly

Leg A (game creation) has real near-term revenue. The market got priced in the last nine months: OpenAI's reported $500M offer for Medal (late 2024, single-source on the figure but the rejected offer is company-confirmed), General Intuition's $2.3B valuation (June 2026), Worldmodeldata's £7M seed (July 2026), Origin Lab's publisher-side marketplace (May 2026), and Grunt Games (Genmo) already paying gamers for video+inputs with per-game bounties.

Leg B (robotics) is a narrative, not evidence. No major robot foundation model trains on game footage (V-JEPA 2, GR00T, Cosmos, π0, Figure, Tesla — verified). The single direct datapoint (D2E, ICLR 2026) is mostly simulation with one thin real-arm win. This matches the Quests as a Data Product buyer note: robotics is a future high-level option (navigation/locomotion priors), never a reason to build manipulation quests. Treat Leg B as a cheap option: keep the capture spec compatible, spend nothing else on it.

How this series fits the existing proposals

The quests thread already encodes the engine's output side: the A–E buyer-value axes, the PREMIUM/COMMODITY triage, the requirements DSL, and cash-per-quest tuned to data value. This series supplies the input side — which gaps to quest, priced how, verified how, sold to whom — and the sequenced plan for proving the whole loop works before scaling it.

Reading order

  1. Market & Buyers — who pays, what's priced, where demand is weakest.
  2. Technical Evidence — why action-aligned data matters; the honest robotics read.
  3. Competition & Moat — the 12–18 month window and what's defensible after it.
  4. Engine Architecture — the flywheel, component by component.
  5. Products & Publishers — the three products and the publisher flywheel.
  6. Risk & Compliance — the two deal-breakers and the non-negotiables.
  7. Roadmap — Phase 0–3 step-by-step, the founding experiment, and the tripwires.