Leash Bio releases Minos, a replayable drug-discovery lab for scoring AI scientist agents

• Minos: an agent searches a library of about 6.3M molecules for the best binders to a protein target, with a budget of about 100 molecules (Leash blog). Latent Space gives the library size as 6.5M (Latent Space).
• The flagship task covers 33 targets. Leash ran about 100,000 trajectories in six weeks and says it adds 2 new environments a day, aiming for 20 a day (Leash blog).
• Release: preprint, open data, Harbor packages and thousands of reasoning traces.
• No language models are named in the blog.

• Evaluation design: the whole library is measured before any agent runs. Every agent decision can then be compared with what the molecules it skipped would have shown. A real lab-in-the-loop run could only learn this by paying for those experiments. Leash puts the number of possible trajectories per target at about 10^638 (Leash blog).
• Inspecting behaviour: stated intent, actual actions and physical outcomes can all be examined, which makes the environment usable for studying agent policies and for reinforcement learning.
• Data controls: DEL molecules carry dozens of barcodes each, to catch false hits caused by protein–DNA binding. Leash also engineered protein mutants ("breaking the pocket") to block model shortcuts (Latent Space).
• Splits: removing shared building blocks between training and test sets cut usable training data by 66% (Owl Posting).

• Agents: stronger models can do the Minos task "at least a little bit." Weaker models stay stuck in local regions of chemical space, leave budget unused, re-test the same molecules and update poorly on contradictory results (Leash blog). No scores or model names are published in the blog, so these findings cannot be checked against rankings.
• Chemist-style leakage: 1,815 authors, 60% top-5 accuracy (arXiv). Owl Posting reports 27% top-1 (Owl Posting).
• Hermes vs Boltz-2: 0.761 vs 0.577 AUROC on Leash's private data. Leash reported this and Owl Posting relayed it; no independent replication was found in the reviewed sources (Owl Posting).
• BELKA: about 2,000 Kaggle teams took part, and none beat chance on unseen chemical libraries (Latent Space).

BELKA was Leash's Kaggle release of 133M molecules screened against three proteins (Introducing BELKA). Its result, near-random performance on unseen building blocks, motivated the strict splits used in Hermes. The Clever Hans paper extends the same leakage argument to public ChEMBL benchmarks. Minos repackages Leash's proprietary DEL measurements as agent environments, which moves the company from benchmark provider to supplier of environments for training and evaluating AI scientists (Latent Space).

• Agents can only choose from molecules that were already screened. Anything outside the measured library has no answer (Leash blog).
• DEL binding signals are not the same as potency, pharmacokinetics or safety.
• Sources differ on figures: library size (6.3M vs 6.5M) and author-ID accuracy (top-5 vs top-1 metrics).
• Owl Posting cautions that chemical space is far larger than any training data, and that Hermes's results need broader validation (Owl Posting).
• No independent evaluation of Minos was found.

• Benchmarks: chemistry benchmarks built from public data should be checked for author leakage. Author-disjoint and building-block-disjoint splits are cheap tests.
• Agent builders: Minos offers a budgeted, verifiable scientific task that scores decision quality, not just final answers. Relevant for anyone training AI-scientist agents with reinforcement learning.
• Strategy: the advantage Leash claims is proprietary, perturbation-designed lab data rather than model architecture (ecosistemastartup). Buyers should still wait for third-party results.