Phase 2 — First Revenue
Convert Phase-1 evidence into signed contracts before the competitive window narrows (Competition & Moat: ~12–18 months). A small signed contract beats a bigger corpus — contracts are the moat; corpus size is optimized last. Phase 2 overlaps Phase 1: outreach starts the moment the Klindt POC writeup exists, without waiting for the Founding Experiment.
Door (a) — One corpus-less world-model lab
Target list, in order
(Revised 2026-07-08 per the vault's grounded buyer knowledge.) Work the named live threads first: Roblox Foundation AI (Q3 refresh; the $350–425K first-batch and $8–10/hr anchors are EC's only real pricing data), FACEIT (revenue-split impasse to revisit with a dollars-per-year case — what Fabio explicitly asked for), and GDM partnerships (rights-clean chain mandatory; their stated openness to joint publisher licensing aligns with our end-state). Then David Klindt's network as the warm research-side path (the POC is the intro artifact — remembering researchers are spec authors, not buyers; they open doors to the procurement side). The world-model-startup lane (Decart, World Labs, Odyssey, Luma, Runway, AMI Labs) is pipeline, not plan-of-record — budget-constrained and synthetic-tempted (Hugo precedent); the founding experiment's directed-vs-synthetic delta is the counter. Apply the GTM plan's "one-no-kills" persona discipline: Research-Scientist wedge, Legal/Procurement veto — the trust packet (provenance manifest) is built for the veto-holder, not the enthusiast.
Pitch assets (all produced by Phase 1)
- The Klindt POC writeup — "your model class, measurably improved by 25 targeted hours."
- The Founding Experiment delta — "directed collection is worth X pp on your weak slices at equal hours."
- The IDM leverage demo — "our labeler multiplies your scraped archive."
- The provenance walk-through — "every clip survives your Art. 53 disclosure" (Phase 0 0.5).
The pitch's spine: we don't sell hours, we sell measured coverage of your eval gaps — the one thing neither Medal (no quests, no raw inputs) nor the publisher marketplaces (no player fleet) can offer.
Contract shape — the early-access arrangement
The template General Intuition proved (reported ~$40M 2025 revenue via early-access arrangements): a paid pilot with expansion structure, not a dataset invoice.
| Element | v1 position |
|---|---|
| Pilot | Fixed-fee directed-collection campaign against 2–3 buyer-specified eval gaps, delivered with measured deltas |
| Expansion | Recurring campaigns; per-verified-coverage-unit pricing calibrated by the experiment's unit-cost number |
| Exclusivity | Slice-level, time-boxed (e.g. exclusive on a title×scenario slice for N months) — never whole-catalogue |
| IP & indemnity | Contributor-rights indemnity only; publisher-layer risk expressly allocated to buyer for non-partnered titles; training-side fair-use bet stays with buyer (Risk & Compliance mitigation 5) |
| Provenance | Machine-readable manifest per delivery; withdrawal propagation terms explicit |
| IDM access | Hosted inference or licensed weights; never the underlying labeled corpus |
Pricing method
Anchor inside the researched band — raw-footage comparables ($60–240/hr of footage) to teleop ceiling ($118–200/hr) — then justify the top of it (or above) with the delta: price per verified coverage unit ≈ unit cost × margin, sanity-checked against what the pp-improvement would cost the buyer in compute/scraping to achieve otherwise. Never quote naked $/hr; the whole strategy is that hours are not the unit.
Door (b) — One studio
Target profile
Indie/AA, PC-first, no kernel anti-cheat, EULA silent-or-friendly on recording, active title with a live-ops need (playtest coverage, balance telemetry). Not EA — the publisher flywheel starts where legal friction is lowest and a single champion can sign.
The motion
Lead with value, not data-rights ask: playtesting agents + player-coverage analytics on their title (Product 3 — the budget line already exists; PlaytestCloud proves it). The quest engine demo is concrete: "we can put 50 real players through your new zone tonight and hand you the coverage map."
The structure — services for rights
An analytics/playtesting engagement whose contract includes AI-training data rights over gameplay of their titles, with revenue share on downstream data products. This converts risk #1 into supply: their title moves to the fully-cleared premium tier (both copyright layers), and the end-state dataset (server-side ground truth + client-side inputs) becomes possible for the first time. Blizzard's separate SC2 AI/ML license is the precedent to cite — publishers already treat ML use as separately licensable; we make granting it profitable.
Publishing strategy
Publish the founding experiment and the naturalness measurement regardless of sign (pre-registered honesty is the credibility asset); publish the IDM cross-game comparison if it beats D2E on KBM titles. The publications are the top of the sales funnel — this market's buyers are researchers.
Supporting workstreams
- Second annealing cycle (Track C5) so the influence scorer stays calibrated as the corpus grows.
- Sales collateral hygiene: every number in the deck traceable to the pre-registered docs — this discipline is the differentiation from data brokers.
- Legal templates (pilot agreement, indemnity schedule, provenance exhibit) reviewed once, reused per deal.
Phase 2 gate
One signed lab contract (pilot or better) and one studio engagement that includes data rights. Both doors, not either — the lab contract prices the engine; the studio engagement starts the moat.
If labs stall but studios bite (or vice versa): proceed on the live door, diagnose the dead one — lab stall usually means the delta isn't compelling (back to Track A) or the rights story isn't clean enough (accelerate door b); studio stall usually means the Product-3 demo is underwhelming (invest in the agents) or the rights ask is landing too early (decouple, earn trust first).