# RL Environment Growth Playbook How a sovereign operator like Navon enters the [[RL Environment Stack]]: lead with custody, not compute. Win the trust layer first, then sell capacity into it. > [!info] Status > Claude's synthesis from the stack analysis. Not sourced. Test each stage against its kill criterion. ## Positioning - Do not fight Prime or Northflank on price per sandbox-hour. That is a commodity race. - Sell what US vendors cannot: **data that legally cannot leave the country**, processed and graded in-region. - [[7 Powers]] fit: cornered resource (custody data), switching costs (provenance chain), counter-positioning (sovereign vs. US cloud). ## Stages ### 0 · Validate (4–6 weeks) - Measure real cost per sandbox-hour on owned hardware vs. the ~$0.039 anchor. - Interview 10 buyers: 5 labs, 5 physical-AI data suppliers. One question: do you pay for custody, or only for price? - **Kill**: no buyer names a custody premium. ### 1 · Wedge: in-region anonymisation - Customer: egocentric and physical-AI data suppliers (vfrog-type). - Offer: face and biometric scrubbing on trusted GPUs, with a signed [[RL Data Provenance]] record. - Why first: a sharp pain, a small GPU footprint, and it fits [[Navon Sovereign Vaults]]. - **Metric**: hours processed per month. 2 paying suppliers. - **Kill**: suppliers stay on marketplace GPUs at a lower price. ### 2 · Distribution: plug into the standards - Ship OpenEnv and Verifiers bindings. List environments on Prime's Environments Hub. - Sell burst CPU sandboxes at or below the anchor, using cheap power. This is the loss leader, not the business. - **Metric**: external trainers running on Navon sandboxes. ### 3 · Expansion: regional domain environments - Build tasks from custody data: Arabic and Swahili workflows, mobile money, public-sector processes, regional regulation. - These are the "next fields" as US-centric tasks saturate. - Partner with governments through [[Data Embassies]]. - **Metric**: environments sold to labs. Re-order rate. ### 4 · Moat: neutral grader hosting - Host and sign graders independently of the environment seller. See [[Reward Integrity]]. - PQC-signed scores and lineage. Long shelf life justifies [[HSM x PQC]]. - **Metric**: third-party environments verified on Navon. ## Flywheel ```mermaid graph LR A[Custody data partners] --> B[Regional environments] B --> C[Lab revenue] C --> D[Trusted grader reputation] D --> A ``` ## Guardrails - One product per stage. Avoid the human-data plus robot-telemetry sprawl. - No single lab above 40% of revenue. - Every task has a retirement date. Plan for ~70% pass-rate decay. ---- tags: #boat #deeptech links: [[Navon MoC]] [[RL Environment Stack]] [[Navon Thesis]] [[Market Intel - Navon]] [[WebSummit Qatar x Regional Go To Market Reflections]] [[Kenya Telco Market x Navon Role]]