Passa al contenuto principale

Phase 3 — Compound

Purpose

Phases 0–2 prove and price the engine; Phase 3 turns the head start into the assets that survive fast-followers (Competition & Moat: contracts, history depth, publisher partnerships, provenance trust — not the capture client). The governing discipline: scale against demand, never ahead of it.

Demand-led fleet scaling​

Quest campaigns are commissioned by contracts; organic capture grows the atlas baseline. Concretely:

  • Fleet growth targets derive from the contract pipeline (committed + weighted-probable campaigns), not from MAU ambitions. The Reward Economy stays solvent by construction — bounty spend is a cost of goods on signed revenue.
  • Contributor quality over quantity: graduated trust (Track C4) concentrates high-bounty work on proven contributors; recruiting targets the catalogue's thin input regimes (KBM-precise FPS players, voice-comfortable narrators, full parties for multi-POV) rather than raw headcount — the N×(N−1) party multiplier makes organized 5-stacks worth more than five solo players.
  • Capacity signal: when contracted campaigns queue beyond ~a sprint of fleet capacity, recruit; when the queue empties, don't.

Publisher partnerships — building the unreplicable dataset​

Door (b) of Phase 2 proved the services-for-rights structure on one studio. Phase 3 systematizes it:

  1. Repeatable package: productized playtesting/analytics offering with a standard rights schedule — each new studio is an integration, not a negotiation from scratch.
  2. The technical prize: server-side ground truth (engine state, authoritative positions) time-aligned with client-side inputs and POV from our fleet — the dataset shape nobody has cross-publisher (Products & Publishers). Each partnered title graduates to the fully-cleared premium tier with engine-state supervision that pixel-only competitors cannot match.
  3. Portfolio target: a meaningful share of active catalogue fully rights-cleared (both copyright layers) — this is the metric that de-fangs risk #1 structurally, not just procedurally.
  4. Publisher revenue share turns rightsholders into stakeholders in the data market's growth — the strongest available deterrent to the enforcement-flip scenario (tripwire T2).

Longitudinal histories — the asset time builds​

Per-player, cross-game records (skill curves, learning trajectories, thousands of hours of context) accrue automatically and cannot be backfilled by late entrants — the one data dimension where the head start compounds instead of eroding.

Productization: expose history depth as a premium slice dimension — "novice→expert learning trajectories on title X," "same player across 5 titles" (cross-embodiment analogue), "10,000-hour veterans encountering a new game." Curriculum-learning and skill-acquisition research buys exactly this, and no archive that started later can offer it. Consent note: longitudinal linkage is a distinct consent scope — handled in the consent architecture (0.3) from day one.

Leg B — re-entry criteria​

Robotics stays a cheap option until evidence moves it (Technical Evidence §C). Re-enter when either:

  1. D2E-class results replicate at scale (multi-task, multi-lab, real-robot deltas from gameplay pretraining), or
  2. A robotics lab publicly buys gameplay data at meaningful scale (tripwire T5).

Until then, the only Leg-B spend remains the title-set bias (first-person, physics-heavy, continuous control) and honest navigation/locomotion-prior labeling on relevant slices. On re-entry: quest design shifts toward the robotics-benchmark style (constrained dynamics, simple goals — the Klindt POC shape at scale), still never manipulation/force quests.

Loop-cadence operations​

The engine's steady state, run as an operational rhythm rather than a project:

  • Per weakness: trigger → quest → collect → retrain → remeasure, ~5–10 rounds (Tesla's pattern), with trigger retirement on saturation (Cruise's anti-bloat rule).
  • Per cycle: annealing recalibration of the influence scorer (Track C5); atlas refresh; naturalness-slice check (the Founding Experiment's step 7 becomes a standing monitor, and its answer sets the organic:directed mixing ratio).
  • Per quarter: title-set and EULA re-audit (Phase 0 0.4 timer); tripwire review against the Roadmap table; pricing recalibration from realized unit costs.
  • The generation-first gate tightens over time: as world models improve, more gaps become synthesizable and the human-quest share shrinks (Engine Architecture §2 — the SIMA 2 trend). Phase 3's posture: welcome it. The durable business is coverage intelligence — knowing what's missing and acquiring it from the cheapest adequate source (fleet, generation, partner) — of which the fleet is one supplier among several that we happen to own.

Metrics​

Publisher partnerships signed; % catalogue rights-cleared (both layers); repeat-buyer rate and net revenue retention on campaigns; fleet retention and cost per verified coverage unit (should fall as the IDM and compiler mature); longitudinal-history depth (median hours/contributor, titles/contributor, history-priced revenue share); organic:directed mixing ratio stability.