RetroEnv-RL

Multistep retrosynthesis tasks for agentic RL. Each task gives a target molecule, a depth budget (longest linear sequence) and optionally a constraint; the agent plans routes with tools and submits synthesis trees whose every leaf must be in the frozen stock. A deterministic verifier judges each step against a frozen reaction library (known reactions and frequent rdchiral retro-templates), never against a hidden answer; known routes only add a similarity bonus.

Variants: max_depth (cap at the shortest known depth while the patent route is longer), restricted_stock (building blocks of both known routes removed), forbidden_class (a reaction class used by both known routes forbidden) and diversity (two or three routes with different first disconnections). Every task ships a witness route proving it is solvable under its constraint.

Held-out targets share no patent, route molecule, reaction, scaffold group or Tanimoto ≥ 0.90 near-duplicate with train or with each other; library reactions touching any held-out key are hidden from the agent's tools (library/reactions.jsonl.gz, visible).

Source: PaRoutes v2 (Genheden & Bjerrum, Zenodo 7341155, CC-BY-4.0), patent reactions from USPTO. Known routes are public, so report results knowing a model may have seen them.