This document records the model audit behind the retrosynthesis Skills. It separates scientific tasks by their input, output, and independently testable hypothesis. Route rendering, de-duplication, provenance, and report generation remain important engineering functions, but they are not counted as scientific model tasks.
A default model must have all of the following:
Paper accuracy alone is not an admission criterion. A model is also rejected as a default when its public artifact only supports training, silently requires a commercial service, or cannot identify the checkpoint that produced a result.
| Independent task | Input | Output | Selected implementation | Engineering verdict |
|---|---|---|---|---|
| Atom mapping and reaction-centre extraction | complete reactant/product reaction SMILES | atom-mapped reaction, mapping confidence, changed bonds | RXNMapper | Mature local package, MIT, pretrained model, CPU/GPU. It cannot map a target-only query because the product and reactant sides must both be known. |
| Single-step precursor generation | one product SMILES | ranked precursor sets for one disconnection | RetroChimera 1 | Preferred proposal model: MIT code and checkpoints, public Pistachio/USPTO checkpoints, direct Python/Syntheseus API. Keep only 5–10 candidates and never interpret its probability as experimental success. |
| Multi-step route search | target SMILES, expansion policy, stock | solved/unsolved route trees | AiZynthFinder | Mature MIT planner with downloadable public policies and stock. The code license does not by itself settle the separately downloaded artifact terms. This is search guided by learned expansion/filter policies, not one end-to-end neural model. |
| Forward outcome and round-trip validation | reactants plus reagents | ranked product SMILES | ReactionT5v2-forward | 2025 peer-reviewed model, MIT 0.2B safetensors on Hugging Face, direct Transformers inference. RetroChimera's released forward checkpoint is a useful same-dataset alternative, not independent evidence. |
| Reaction-condition recommendation | full reaction SMILES | ranked catalyst/reagent/solvent classes | Parrot USPTO | First-author HF revision b9ef604... explicitly declares MIT; both artifacts are size/hash pinned and a real GPU worker canary returned 15 joint beams. The legacy Google Drive artifacts remain blocked. This checkpoint emits categorical labels and does not support temperature; frozen benchmark accuracy is still pending. |
| Reaction-yield estimation | reactants, reagents, and product | predicted isolated yield percentage | ReactionT5v2-yield (quarantined release) | The 2025 MIT checkpoint loads locally, but the pinned release failed its published canary after upstream preprocessing was reproduced (about 19.1666 expected, 65.924858 observed). Keep it protocol-only until independently resolved and validated. |
These six tasks form a useful model-backed planning and review surface. Parrot is admitted only for the exact first-author HF snapshot recorded in the model manifest; any other checkpoint still requires a separate pre-download allow decision. These tasks do not make a synthesis experimentally complete. Procurement, EHS review, scale-up, work-up, purification, analytical release, and literature/ELN verification still require databases, rules, experiments, and chemists.
The tasks can therefore be trained and evaluated separately. Their outputs also must not be collapsed into one confidence number: mapping confidence, beam likelihood, search score, product rank, condition top-k accuracy, and yield regression error have different meanings.
aizynthmodels; new deployments should not
start from the archived Python 3.7 stack.Sources were last rechecked on 2026-08-19. Pin repository revisions and model artifact hashes in any regulated or long-lived deployment; a moving model name is not provenance.