# Sellable RL Environment An AI lab buys an RL environment when three parts come together: data only you hold, the workflow that gives it meaning, and a grader that checks the answer. - Data alone is a dataset. Labs can get datasets, often for free. - The workflow turns data into a job a model can practise. - The grader tells the model when it did the job right. Labs pay for this score. ```mermaid graph TD A[Data only you hold] --> D[RL environment] B[Expert workflow] --> D C[Grader] --> D D --> E[Lab trains a model on a real job] ``` ## The three parts Example (illustrative): a crop insurer in East Africa. | Part | Plain meaning | Crop insurer | |---|---|---| | **Data** | Records nobody else has | 10 years of claims, field photos, weather, payouts | | **Workflow** | The steps an expert takes | Check the policy, read the photos, compare to rainfall, decide the payout | | **Grader** | A check a machine can run | Does the model's payout match the settled claim? | ## What breaks if a part is missing - **No unique data.** Labs build it in-house or use open data. Build AI released Egocentric-1M free. Free data caps the price. - **No workflow.** You sell raw hours. Buyers want proof that the model gets better at a job, not volume. - **Weak grader.** The model learns to game the score. See [[Reward Integrity]]. ## How a domain operator gets ready 1. Find decisions with a known answer later: a settled claim, an approved loan, a passed inspection. 2. Freeze the data with dates and versions. Train on old records. Test on newer ones. 3. Write the expert steps as tasks. Give the model the same tools the expert uses. 4. Write the grader. Try to cheat it before you sell it. 5. Run current models. Keep the tasks they fail. Drop tasks they pass at ~70%. 6. Record consent and lineage for each record. See [[RL Data Provenance]]. 7. Publish a small open set as a storefront. Keep the held-out set private and paid. ## Why local operators have an edge - They hold data that cannot leave the country. - They know languages and workflows that labs lack: Arabic and Swahili workflows, mobile money, public-sector processes, regional regulation. - Their experts know what "correct" means. See [[Domain Experts as Eval Builders]]. - Labs need new fields as common tasks saturate. ## Who buys | Buyer | What they want | |---|---| | Frontier labs | New fields where models still fail | | Neo labs | Depth in one vertical, such as biology or law | | Physical-AI labs | Real-world sensor and video data, with consent records | ## Examples - **Evos.** Open-sourced 8 agent skills for logistics, manufacturing and energy. Each ships with 20–25 real scenarios, scored by rubrics. This is workflow plus grader from ops veterans. See [[Agent Skills as Codified Domain Expertise]]. - **Versioned biology data.** Train on database version N. Test on facts that first appear in N+1. The ground truth becomes the grader. See [[Drugverse RL Environments and Data Play]]. - **Prime Intellect Environments Hub.** 2,500+ listed environments. A storefront where a domain seller can list an open set. - **Counter-example: Build AI.** Open video data with no workflow or grader. It sets a low price for raw hours. Technical layers: [[RL Environment Stack]]. ---- tags: #boat #deeptech links: [[Foundational Models MOC]] [[RL Environment Stack]] [[Data Moat]] [[Domain Experts as Eval Builders]] [[Evaluation and Verification Infrastructure]]