The position: Playroll is not a data brokerage and not a frontier lab. It is a data engine — a system that knows where world models are weak, sends players to fill exactly those gaps, proves the improvement with pre-registered benchmarks, and sells the measured result.
OpenAI reportedly offered $500M for Medal; the spin-out hit $2.3B. Marketplaces for game data raised seeds this month. Our thesis is validated — and the clock is running.
Everywhere it's been measured, directed collection beats passive volume (Microsoft: 1 curated week ≈ 7 passive years; Waymo: +31% from 3% targeted). Nobody has proven it for game world models. We will be first.
Our audit of all 19 titles: 0 green · 4 amber · 15 red. Selling footage is contractually barred nearly everywhere. So we sell measured improvements and services — and make publishers partners, not plaintiffs.
Pre-registered benchmarks, frozen metrics, honest negatives. The whole commercial story reduces to “our numbers are real.” The first massaged threshold ends the business — protect this above all.
Nine months ago there was no market for action-aligned gameplay data. Now it is being entered from three directions at once: player fleets, publisher marketplaces, and input-reconstruction from free video.
Reported $500M acquisition offer for gameplay clips. Rejected.
Medal spins out a lab: $133.7M seed → $2.3B Series A (Jun ’26).
$8M seed — publishers sell rights-cleared game data to labs.
£7M seed — the competitor David flagged. Targets 1M licensed hours.
Genmo pays gamers for video + inputs, per-game bounties. Our recipe, replicated.
Our window: ~12–18 months of being the only supplier of consented, input-precise, directable sessions — before Medal reverses its no-raw-inputs stance or Overwolf bolts quests onto 113M users. The moat that survives after that is built from contracts, publisher partnerships, longitudinal player histories, and trust — not the capture client.
Every arrow is colored by how much evidence stands behind it today.
2K labeled hours unlocked 70K scraped hours for OpenAI (35:1 leverage). Every playable world model ever built needed action labels. Ours are real, ms-aligned, cross-game.
Our external research, the team’s June strategy work, and David Klindt’s own proposal all landed on the same thesis independently: directed, benchmark-first, coverage-audited capture.
Directed vs passive data, equal hours, pre-registered slices. Costs ~$8–20K per arm at pilot scale. A positive delta is our pricing evidence, pitch deck, and publication — in one number.
The plan was checked against every prior doc, 99 days of operational analytics, and the full vault. What survived, survived scrutiny. What didn’t is corrected below — in the docs, not just here.
A whitelist of “green” titles exists for selling gameplay footage.
Zero green. Five publishers added explicit AI/TDM bans in 2024–26; five claim ownership of recordings outright. Amber set for internal R&D: Minecraft, BG3, Elden Ring. Everything else routes through partnership.
Corpus-less world-model startups are the primary buyers.
Procurement happens at frontier labs and publisher AI programs. Our live threads: Roblox Foundation AI ($350–425K first-batch anchor), FACEIT, GDM (open to joint publisher licensing). Startups are pipeline, not plan.
A fleet exists to collect hundreds of hours per experiment arm.
1–4 people record per day; ~100 gross hours ever. The POC (20–30h ≈ 3–4 operator-days) is sized right. Anything bigger is a funding decision, not a scheduling one.
“Mic-only, never incoming voice” is a policy line.
The pipeline captures game-mixed voice chat today (and once caught a dev’s Discord). This is the one act-this-week item: quarantine existing game-audio tracks, ship isolation or track-drop, counsel review.
The coverage system is new infrastructure to design.
The team designed its world-space half in June (realtime-spatial-coverage). Merged, not parallel-built — and made learnability-weighted per David’s “broken TV” caveat: reward what improves the model, not raw novelty.
Gates, not dates. Each phase earns the next.
18+ verification · consent scopes for AI/sale · title tiers · provenance ledger · audio fix.
Minecraft, ~20–30h, pre-registered dynamics probes (persistence, permanence, inventory… + geometry). Fine-tune, measure, return to David with data and his asked-for format.
Directed vs passive at equal hours. Pilot 50h/arm is feasible now; full scale is a funded operator program.
One lab/publisher-AI contract (Roblox/FACEIT/GDM threads first) + one studio deal structured to include data rights.
Publisher partnerships → the dataset nobody else can build (server truth + client inputs). Histories accrue; trust compounds.
The deeper play: the durable business isn’t the fleet — it’s coverage intelligence: knowing what world models are missing and acquiring it from the cheapest adequate source. The fleet is one supplier of that. We happen to own it.
| Owner | Action | Unblocks |
|---|---|---|
| Everyone | Treat the audio quarantine as the only “this week” item | Legal exposure that already exists |
| Antonio | Pick probes from the B-04 survey; 5-min MineWorld download check | POC pre-registration + model choice |
| Vincenzo | Corpus-sample cut (with hour counts, consent-era filter, audio handling) | Atlas, curation, IDM briefs |
| Alex | Counsel packet out (cover note drafted); ingest the 07-07 call into the vault | Rights posture; team knowledge |
| Marco / infra | Paste the two session handoffs (eval harness on dfgpu; provenance schema on Supabase staging) | The POC’s critical path |
| Phil | Operator capacity for ~3–4 Minecraft days, scheduled around the offsite | Collection |
Everything above lives in full depth in the docs: Proposals → World-Model Data Engine (10 pages + 7 roadmap sub-pages + the AI delegation loop), with returned research artifacts in ec-workspace/research/world-model-data-flywheel/.