# Drugverse RL Environments and Data Play Drugverse already owns the hard part of a biology RL environment: integrated, versioned, ID-linked ground truth. Ground truth gives verifiable rewards. Labs pay for verifiable rewards. > [!info] Status > Strategy synthesis, Oct 2026. Not tested with buyers. Test each stage against its kill criterion. ## Why Drugverse fits - RL needs tasks with answers a machine can check. See [[RL Environment Stack]]. - Public bio databases have the answers, but in pieces. [[Drugverse]] joins the pieces by ID. - **Versioning is the edge.** Train on source version N. Test on facts that first appear in version N+1. This gives a held-out set with no leakage. Few vendors can do this. ## What to sell | SKU | What | Buyer | |---|---|---| | **Environments** | Task sets + graders + tool harness, built on Verifiers | Neo labs, frontier lab bio teams | | **Private evals** | Held-out, time-split benchmarks. Signed scores | Same, plus pharma AI teams | | **Data licences** | Integrated snapshots, version diffs, agent trajectories (SFT traces) | Model builders | ## Task families - **Entity resolution.** "What is the current HGNC symbol for X?" Reward: exact ID match. - **Cross-reference.** Map a DrugBank ID to IDs in other sources. Reward: set F1. - **Drug → target → mechanism.** "Which targets does drug X act on?" Reward: set overlap with ground truth. - **Repurposing (multi-hop).** "Which approved drugs hit genes linked to disease Y?" Reward: ranked recall against the graph. - **Agentic tool use.** The model gets read-only SQL and graph tools over a frozen Postgres snapshot. Multi-turn. Reward: correct answer, then fewer tool calls. - **Time-split forecast.** "Which new target–disease links will appear?" Reward: match against the next source version. ## How to build on Prime Intellect 1. Install `verifiers`. Pin the version. The v0.1.14 release (May 2026) changed the API to Taskset/Harness.^[Source: [PyPI: verifiers](https://pypi.org/project/verifiers/)] 2. Build the dataset from a frozen Postgres snapshot. Store prompt, answer, source IDs and source version per row. 3. Write the rubric: one reward function per task family. 4. Write the harness: single-turn for QA, multi-turn with tools for agentic tasks. 5. Run baselines on 3–5 open and frontier models. Drop tasks that models pass at ≥70%. See [[Reward Integrity]]. 6. Publish a **small public subset** to the Environments Hub as a storefront. Keep held-out sets private and paid. 7. Add an OpenEnv binding later, for buyers that do not use Verifiers. Docs: [verifiers repo](https://github.com/PrimeIntellect-ai/verifiers). Check the current publish commands there. ## Go-to-market ### 0 · Validate (4 weeks) - Talk to 5 buyers from [[Biotech Neo Labs - Buyer List]]. - Ask: "Where do your models fail in biology? Do you pay for held-out, versioned bio evals?" - **Kill**: no buyer names a gap that Drugverse data covers. ### 1 · Storefront - 2–3 open environments on the Hub. Publish a leaderboard of frontier models. - **Metric**: external runs, inbound lab contacts. ### 2 · Paid pilots - Private eval + environment contract with 1–2 neo labs. - Price anchor: $300–500k per quarter for smaller labs.^[Source: [Troveo: How to sell RL environments](https://www.troveo.ai/resources/sell-rl-environments). Vendor source, anecdotal.] - **Kill**: labs take the open set and build the rest in-house. ### 3 · Data foundry - Each new source version makes new tasks automatically. This answers alpha decay. - **Metric**: re-order rate. ## Risks - **Source licences.** DrugBank needs a commercial licence. Check the licence of every source before a sale. Sell only what the licence allows, or sell graders and tasks without raw data. - Labs build bio environments in-house. OpenAI builds its own data team.^[Source: [SemiAnalysis: RL environments and RL for science](https://newsletter.semianalysis.com/p/rl-environments-and-rl-for-science)] - No public evidence yet of labs buying bio-specific environments. This is a bet on a new category. - Reward hacking on loose graders. ---- tags: #boat #deeptech links: [[Drugverse]] [[TechBio MoC]] [[RL Environment Stack]] [[RL Environment Growth Playbook]] [[Reward Integrity]] [[RL Data Provenance]] [[Biotech Neo Labs - Buyer List]]