Roadmap & Playbook
Expands Deliverable 6 §5–8 of the research pass into an executable playbook. Durations are planning anchors, not promises; gates, not dates, control phase transitions. The competitive clock that motivates the pace: ~12–18 months before Medal can credibly reverse its no-raw-keys stance or Overwolf bolts quests onto 113M MAU (Competition & Moat).
This page is the overview; each phase and track has a full sub-page: Phase 0 · Track A0 — Klindt POC · Track A — Founding Experiment · Track B — Generalist-IDM · Track C — Engine Plumbing · Phase 2 · Phase 3. Track A0 was added after the David Klindt call + team debrief (2026-07-07), which independently converged on this plan's Phase 1 and sharpened its first step.
Alignment review 2026-07-08 (ec-workspace/research/world-model-data-flywheel/alignment-review-2026-07-08.md):
this plan was checked against the full prior docs tree and operational
analytics. Validated: directability-as-moat is EC's own prior conclusion
(realtime-spatial-coverage); the POC's RCON
mechanism is already built. Corrected here: capture spec is 1080p60 masters
(not 720p). Still open (team decisions): catalog↔EULA-audit reconciliation;
founding-experiment scale vs real fleet throughput (~1–4 recording
contributors/day — full N requires a dedicated operator program); voice
strategy (shipped promise is transcript-only); bug-14/EC-113 audio
contamination; bug-16-era consent records; coverage-unit basis type in the
payments schema. Read the review before executing Phase 0/1 items.
The shape of the plan
Ordering logic: legal posture before collection (nothing gathered under a weak posture is sellable later); proof before scale (the founding experiment is cheapest when the fleet is small); revenue before corpus (contracts are the moat, corpus size is optimized last).
Phase 0 — Foundations (now → +1 quarter)
Full detail: Phase 0 — Foundations.
Everything in this phase exists so that every hour collected afterward is sellable. The Risk & Compliance non-negotiables are the checklist; the steps below sequence them.
| # | Step | Detail | Done when |
|---|---|---|---|
| 0.1 | Lock the capture spec | ms-aligned K/M/controller streams; 1080p60 masters (recorder already enforces ≥1080p on the quest lane; the 3D-reconstruction leg needs the fidelity — 720p30 as derived working copies only); mic-only audio, incoming voice channels never captured (⚠ current pipeline violates this — bug-14 loopback + EC-113 game-mixed voice; see alignment review); session metadata; per-clip consent reference. Freeze it — VPT locked settings for consistency and so do we | Spec doc merged; client enforces it; changes require ADR |
| 0.2 | 18+ KYC-grade age verification | Real age assurance at contributor onboarding, not a checkbox. Payment rails (bank/PayPal KYC) can carry part of the load | No payout possible to an unverified contributor |
| 0.3 | Consent v-next | Granular, revocable, plain-language: video / inputs / mic / metadata / sale / resale / model classes. Builds on the shipped consent-versioning architecture; SAG-AFTRA 2025 IMA as the drafting benchmark | Counsel-reviewed; versioned; re-consent flow tested |
| 0.4 | Title whitelist + EULA audit | Per-title review with dated EULA snapshots. Hard blacklist: Nintendo, Frontier-style AI clauses. Flag Valve titles as no-sale-purpose pending partnership. Cross-check against the live catalogue in res/games.json (Game Catalog) | Every enabled title has a dated clearance record and a risk tier |
| 0.5 | Provenance ledger v1 | Per-clip: contributor consent version, title whitelist status at capture, processing lineage; withdrawal/redaction implemented (Ego4D-grade) | A sample dataset export ships with a machine-readable provenance manifest |
| 0.6 | PII pipeline v1 | Blur/strip usernames, chat panes, friend lists; quest-authoring rule: never elicit on-screen PII | Runs in the processing pipeline; spot-check QA process defined |
| 0.7 | Anti-cheat posture review | No hooks in protected processes; per-title technical vetting; open the whitelisting conversation with anti-cheat vendors (Overwolf took years — start now) | Vetting checklist per title; first vendor contact made |
| 0.8 | Pick the Phase-1 title set | 3–5 whitelisted, first-person, physics-heavy, continuous-control titles (keeps Leg-B optionality free), biased to KBM precision and voice-intent (the two data assets reconstruction can't touch). Use the data-domain table — the driving/traversal/FPS-squad domains dominate | Title set frozen for the experiment |
Phase 0 gate: an auditor (or a buyer's counsel) could trace any clip to a consenting, age-verified adult, on a cleared title, under a stated consent version — with withdrawal honored.
Phase 1 — Prove the engine (+1 → +3 quarters)
Track A0 runs first and fast; then three tracks in parallel, with the founding experiment as the headline deliverable.
Track A0 — The Klindt POC (benchmark-first, ~20–30 h, Minecraft)
Full detail: Track A0 — Klindt POC. Source: David Klindt call + debrief, 2026-07-07.
A data-value demo that precedes and de-risks the founding experiment: pick an open Minecraft world model (open Oasis weights / MineWorld class), pre-register a small dynamics benchmark where it demonstrably fails (benchmark before samples — the debrief's resolution), collect ~20–30 operator hours via two minimal-scaffolding quests (construction / destruction-collection), fine-tune, re-measure, and return to David with the writeup. Minecraft is deliberately testbed, not product — it's the one game where open action-labeled data is already commodity; we prove the method there and sell it on the cross-game catalogue. Optional rider: the overlay-narration friction test (intent data, axis A).
Track A — The founding experiment (the company's first evidence)
Full detail: Track A — Founding Experiment.
Claim under test: at equal hours and equal compute, coverage-directed (quest-driven) fleet data beats passive fleet data for world-model fine-tuning — with the largest deltas on rare/OOD slices.
Step by step:
- Build the coverage atlas v1 over the Phase-1 titles: embed the existing corpus (video+action windows), fit a density model, define cells; identify the thinnest cells (Engine Architecture §1). Prior art predicts the payoff: Waymo got +31% on rare slices from 3% targeted additions.
- Define held-out eval slices before any collection: per-title common slices, rare-cell slices, an OOD slice (held-out title), and the naturalness slice (directed vs organic distribution comparison — see step 7).
- Author the directed arm with quest compiler v1: novelty-engine prototype pointed at the thin cells; quests triaged by the A–E axes; many phrasings per quest; per-contributor caps; bounties per verified coverage unit priced via the Reward Economy spine.
- Collect both arms concurrently, same fleet, same weeks (controls for contributor mix and meta shifts): N hours passive/organic, N hours directed. N sized by what an open model fine-tune can detect — start ~200– 500 h/arm and let power analysis on pilot deltas set the final N.
- Train matched fine-tunes of an open world model (Matrix-Game 2.0/ MineWorld class) — identical base checkpoint, compute, and recipe; only the data arm differs. Add a mixed arm (50/50) — production will always blend, and the mix result is the sellable configuration.
- Evaluate on the pre-registered slices. Success = directed ≥ passive overall AND directed ≫ passive on rare/OOD slices. Pre-register internally; publish either way.
- Measure quest-induced distribution shift on the naturalness slice. Nobody has measured this for gameplay; publishing it (either way) is cheap credibility and de-risks the product.
- Write the two numbers on the wall: pp-delta on rare slices at equal hours, and cost per verified coverage unit. These price Product 1.
Track B — Generalist-IDM v1
Full detail: Track B — Generalist-IDM.
- Train an IDM on the labeled corpus across the Phase-1 titles (VPT recipe, non-causal).
- Milestone: label held-out scraped video of an unseen title at usable accuracy — the cross-game bar D2E set, targeted specifically at KBM titles where overlay-parsing (NitroGen) can't compete.
- Deliverable: leverage demo — "our N labeled hours unlock M scraped hours" with measured downstream quality. This is Product 2's proof artifact.
Track C — Engine plumbing
Full detail: Track C — Engine Plumbing.
Quest compiler v1 (gap list → goal + verifier → requirements DSL per
Quests — Technical Design); trigger library +
coverage signatures in the overlay; curation pipeline (dedup/filter/taxonomy);
gold quests + fraud detection v1; distilled influence scorer v1 (MATES
pattern) validated against one annealing ablation (Llama 3 pattern).
Phase 1 gate: the Klindt POC writeup is in David's hands with at least one pre-registered probe showing a clear base→fine-tuned improvement; the experiment shows a delta worth putting in front of a buyer; and the IDM shows cross-game transfer. If the POC shows no improvement → interrogate benchmark design before the thesis (see the POC failure reading). If the experiment shows no directed-vs-passive delta → tripwire T4 below.
Phase 2 — First revenue (+2 → +4 quarters, overlapping Phase 1)
Full detail: Phase 2 — First Revenue.
Two doors; walk through both in parallel:
| Door | Target | Motion | Contract shape |
|---|---|---|---|
| (a) One corpus-less world-model lab | Decart, World Labs, Odyssey, Luma, Runway, AMI Labs class (Market & Buyers) | Pitch = the founding-experiment delta + the IDM leverage demo + provenance ledger walk-through. Offer a directed-collection pilot on their eval gaps | Early-access arrangement (the General Intuition template): paid pilot → recurring campaigns; exclusivity on specific slices in exchange for commitment; indemnity scoped to contributor claims only, training-side risk stays with buyer |
| (b) One studio | Indie/AA, PC-first, no kernel anti-cheat, EULA silent-or-friendly | Playtesting agents + player-coverage analytics on their title; structured to include AI-training data rights (the publisher flywheel entry point) | Services + rights: analytics engagement whose contract grants gameplay-data rights, with revenue share on downstream data products |
Supporting steps: pricing collateral from the experiment numbers (per verified coverage unit vs the $60–240/hr footage band); publish the experiment + naturalness result (the marketing is the science); begin the second annealing cycle so the influence scorer stays calibrated as the corpus grows.
Phase 2 gate: one signed lab contract and one studio engagement with data rights. A small signed contract beats a bigger corpus.
Phase 3 — Compound (+4 quarters →)
Full detail: Phase 3 — Compound.
- Scale the fleet against demand, never ahead of it. Quest campaigns are commissioned by contracts; organic capture grows the atlas baseline.
- Expand publisher partnerships toward the server-truth + client-inputs dataset — the 3-year moat candidate. Target: meaningful share of the catalogue fully rights-cleared (both copyright layers).
- Let longitudinal histories accrue — per-player cross-game records are the asset late entrants cannot backfill; surface them as a premium slice dimension (skill curves, learning trajectories).
- Revisit Leg B (robotics) only on evidence: D2E-class results replicating at scale, or a robotics lab publicly buying gameplay data. Until then the only Leg-B spend is the title-selection bias from Phase 0.8.
- Iterate the loop cadence: trigger → quest → retrain → remeasure per weakness (Tesla ran ~7 rounds per network); retire solved cells (Cruise's anti-bloat rule).
Tripwires — what changes this plan
| # | Signal | Response |
|---|---|---|
| T1 | Medal launches an opt-in raw-input tier | Accelerate contract-signing; differentiation narrows to directability + provenance + KBM precision + histories — lean on all four |
| T2 | A publisher sues any gameplay-data seller | Freeze non-partnered collection immediately; the publisher-partnership track becomes the whole company |
| T3 | Input reconstruction reaches KBM precision | Labeled-input scarcity collapses; value concentrates in directability, voice-intent, and histories — the flywheel matters more, the archive less |
| T4 | Founding experiment shows no directed-vs-passive delta | The flywheel is a nice-to-have: pivot weight to publisher-partnered capture + Product 3 (playtesting), where the fleet has value without the coverage thesis |
| T5 | A robotics lab publicly buys gameplay data at scale | Leg B moves from option to roadmap; bias title selection and quest design accordingly (still no manipulation quests) |
| T6 | COPPA 2.0 passes the House / age-verification mandates land | Already compliant if Phase 0.2 held — audit and certify, turn it into a sales asset |
Metrics per phase
| Phase | Metrics |
|---|---|
| 0 | % of enabled titles with dated clearance records; consent v-next adoption; capture-spec conformance |
| 1 | Experiment delta (pp on rare slices at equal hours); IDM cross-game accuracy; cost per verified coverage unit; naturalness gap measured & published |
| 2 | Signed lab contract (count, $); studio engagement with data rights; $ per delivered coverage unit vs the footage band; repeat-campaign rate |
| 3 | Publisher partnerships; % catalogue rights-cleared (both layers); repeat-buyer rate; fleet retention; longitudinal-history depth (median hrs/contributor, titles/contributor) |