The Athanor ML Family

What the family is

arbastro Intelligence Athanor is the family of chemistry models inside Athanor, our chemistry drawing app. They plan retrosyntheses, predict reaction products and conditions, propose mechanisms, recognize structures from images and handwriting, and predict NMR chemical shifts. This report covers their architecture, training data, evaluation results and deployment status as of November 12, 2026, and includes experimental models alongside the ones running in production.

Athanor ran on a general-purpose language model to begin with, and for product prediction that mostly worked. It was expensive on every request, it slowed down as structures grew, and there was one question we could not answer: we had no way to know whether a benchmark we wanted to quote was already in that model's training data. Training our own settled that. The models are small enough that an interactive prediction costs no API fee, we hold the training set and can audit a result that looks too good, and corrections made in the app can feed the next version.

Three results below have implications beyond their scores. Single-step accuracy did not translate into better route planning. Synthetic training images did not generalize to real patent drawings. Accurate predictions did not come with usable uncertainty.

Family directory

Training status, evaluation outcome and deployment status are tracked separately, because several models here finished training and passed their benchmarks without being deployed.

Model generations use Roman numerals. Experimental checkpoint identifiers and API versions are listed in the model cards.

Every benchmark, at a glance

Every measurement in this report, one chart each, with the deployed model marked and the highest score in each chart labeled. Three of our systems have been evaluated against external tools, across ten of these charts; the remaining charts compare our own versions or a baseline we implemented. The filter selects between them.

How the pieces fit

Several requests are answered by software and rules rather than by a trained model. Route planning combines the single-step policy with Texere, our search engine, and a lookup of purchasable compounds. Mechanisms pass through an atom mapper, a set of textbook rules, the learned model where no rule matches, and an execution engine that validates each arrow. Chemical naming uses OPSIN^{35[35]Lowe, D. M. OPSIN: Open Parser for Systematic IUPAC Nomenclature.}, and structure handling throughout uses RDKit^{33[33]RDKit: open-source cheminformatics.}.

This matters when reading the benchmarks: an improvement in the mechanism service can come from the rules rather than the model, and an improvement in route recovery can come from the search rather than the policy. Each section states which component a result measures.

• Texere (our software). The route search: a best-first search over an and-or tree of molecules and reactions.
• Mapper (our software). Atom mapping by template reversal, with a maximum-common-substructure fallback.
• The mechanism rules (our rules). The same textbook rules that write Motus’s training data, run ahead of the model at serving time.
• Stock lookup (our software and public data). Whether a molecule is purchasable, from public vendor deposits.
• OPSIN (third party). Chemical names to structures, and the round-trip check behind naming.
• RDKit (third party). Structure handling, depiction, templates and the scoring used throughout.
• Quantum calculations (our service). Infrared spectra and orbitals, computed rather than predicted by a trained model.
• athanor-imago-1 (an alias). Cernere I under its pre-rename identifier. Not a separate model.

Data, licensing and leakage control

Every reaction model here is trained on the same source: reactions text-mined from United States patents between 1976 and 2016, released to the public domain by Lowe^{1[1]Lowe, D. M. (2012) Extraction of chemical structures and reactions from the literature. PhD thesis, University of Cambridge. doi:10.6084/m9.figshare.5104873.}. Patent reaction data is weighted toward medicinal and process chemistry, and predominantly describes successful reactions, so it provides little information about what does not work.

The corpus repeats itself heavily: the same reaction appears in a grant, its application, and every related filing. We deduplicate on the reactants and products, re-map every atom ourselves rather than trusting the source's mapping, and then split by patent family rather than by reaction, so that two filings describing the same chemistry cannot land on opposite sides of the split. Before splitting, every reaction that appears in any public benchmark we report against is removed from training, matched by structure in several ways and then checked independently.

Deduplication removes 62.1% of the extracted rows. Benchmark exclusion removes a further 146,515, of which 200 were identified only by connectivity key rather than by structure match. Two patents join a group when they share at least 5 reactions, giving 181,464 groups, the largest holding 0.1% of rows.

Independent verification over 1,150,519 training rows found 0 violations of the exclusion criteria. It reports two conditions it does not treat as violations: 12,653 of 214,071 patents are cited on reactions in more than one split, and 280 of 5,007 products in one public test set are also synthesized in the training data from different starting materials.

Route planning depends on one further dataset: the index of purchasable compounds that terminates a search. 56,394,857 compounds from public vendor deposits^{4[4]Kim, S. et al. (2023) PubChem 2023 update. Nucleic Acids Research.}. The vendor-count threshold changes the index size by roughly a factor of two, so the policy used here is stated below.

The remaining models use openly licensed datasets: published mechanism labels^{5[5]Chen, S., Babazade, R., Kim, T., Han, S., Jung, Y. (2024) A large-scale reaction dataset of mechanistic pathways of organic reactions. Scientific Data 11, 863. doi:10.6084/m9.figshare.24797220.}, real patent drawings from USPTO-30K^{6[6]Morin, L. et al. (2023) MolGrapher: graph-based visual recognition of chemical structures.}, experimental NMR spectra^{7[7]Wang, Z., Zhu, X. (2025) NMRexp: an experimental NMR spectral dataset. doi:10.5281/zenodo.17296666.} and experimental pKa values^{8[8]Zdrazil, B. et al. (2024) The ChEMBL database in 2023. Nucleic Acids Research 52(D1).}. Two widely used online-handwriting corpora and one drawing dataset were excluded on licensing grounds, so handwriting recognition is trained and evaluated on synthetic ink only.

How to read the numbers

Every comparison on this page belongs to one of four classes, and each figure says which. A direct comparison runs another tool on the same examples with the same scorer. A baseline is an algorithm we implemented and ran under the same protocol. An internal comparison is between our own versions. Published context uses a different protocol and is never drawn on the same chart as our numbers.

Results that were never measured are labeled "not measured"; a zero is a measured zero. Intervals are labeled with their type, and where a source records only a standard error, that is what is shown.

Every benchmark on one table

The whole family in one place, a tab per task area, with our result beside whatever it was actually measured against. Each table downloads as a file, and the notes under it carry the caveats a number needs to be read correctly.

These benchmarks measure agreement with a recorded answer. A route that is cheaper or safer than the one in the patent scores zero, and a structure read with the wrong salt form counts as an error.

Retrosynthesis

Retrosynthetic planning separates into two components: a single-step model that proposes disconnections of a target molecule, and a search that assembles those steps into a route terminating in purchasable compounds. Route quality depends on both.

The whole route

Texere is a best-first search over an and-or tree of molecules and reactions, following Retro*^{21[21]Chen, B., Li, C., Dai, H., Song, L. (2020) Retro*: learning retrosynthetic planning with neural guided A* search. ICML 2020.}, with half of its expansion budget reserved for keeping alternative first disconnections alive. We evaluated it on PaRoutes^{3[3]Genheden, S., Bjerrum, E. (2022) PaRoutes: towards a framework for benchmarking retrosynthesis route predictions. Digital Discovery. doi:10.5281/zenodo.6275421.} against AiZynthFinder^{22[22]Genheden, S. et al. (2020) AiZynthFinder: a fast, robust and flexible open-source software for retrosynthetic planning. Journal of Cheminformatics 12.}, which we ran ourselves using its published settings, on the same targets and the same stock.

On the 10,000 targets of the n1 set, Texere solved 98.0% against 96.7% and ranked the recorded route first for 27.1% against 20.0%. AiZynthFinder recovered the recorded route within its top ten more often: 60.4% against 53.6%.

Resolvere I

Resolvere ranks reaction templates from a product fingerprint, following the expansion-policy design of Segler and Waller^{9[9]Segler, M. H. S., Waller, M. P. (2017) Neural-symbolic machine learning for retrosynthesis and reaction prediction. Chemistry: A European Journal 23(25).}. At 79.7M parameters a full training run takes minutes.

A template model can only propose transformations its library contains, so the proportion of test reactions covered by the library sets a hard ceiling on accuracy. The specific library reached 62.3% coverage, and the model retrieved most of what that allowed. Generalizing the templates raised the ceiling to 73.2% and top-50 accuracy to 66.2%, while top-1 accuracy fell from 26.0% to 24.7%, as more near-duplicate templates compete for the first rank. Version 1.1 is deployed.

Rescoring the policy's top templates with the forward model raised top-1 accuracy by about one point on the same reactions. This remains an experiment and is not part of the deployed policy.

Resolvere Liber

Liber predicts precursor structures directly with an encoder-decoder Transformer^{17[17]Vaswani, A. et al. (2017) Attention is all you need. NeurIPS 2017.}, requiring no template library, in the line of Liu and colleagues^{12[12]Liu, B. et al. (2017) Retrosynthetic reaction prediction using neural sequence-to-sequence models. ACS Central Science 3(10).}. At its best-scored checkpoint, 296,000 steps, it reached 32.7% top-1 against Resolvere's 26.5%. A second run using root-aligned representations^{13[13]Zhong, Z. et al. (2022) Root-aligned SMILES: a tight representation for chemical reaction prediction. Chemical Science.} converged faster early in training with no advantage at the end.

Integrated at the first disconnection of the search over 500 targets, top-1 route recovery fell from 25.2% to 21.8%, top-5 and top-10 recovery improved, and median search time rose from 15.5 to 44.2 seconds. Liber is not deployed. The experiment was preregistered on the larger n5 set with 1,000 targets and run on n1 with 500.

Ordo I and Probatio I

Two further models were trained to improve what the search returns. Ordo ranks completed routes using a linear scorer over 11 route features, fitted on 871 route pairs labeled by a language model against a written rubric drawn from the route-quality literature^{26[26]Wender, P. A., Verma, V. A., Paxton, T. J., Pillow, T. H. (2008) Function-oriented synthesis, step economy, and drug design. Accounts of Chemical Research 41(1).,27[27]Burns, N. Z., Baran, P. S., Hoffmann, R. W. (2009) Redox economy in organic synthesis. Angewandte Chemie 48(16).,28[28]Badowski, T., Molga, K., Grzybowski, B. A. (2019) Selection of cost-effective yet chemically diverse pathways from the networks of computer-generated retrosynthetic plans. Chemical Science 10.,29[29]Butters, M. et al. (2006) Critical assessment of pharmaceutical processes: a rationale for changing the synthetic route. Chemical Reviews 106(7).}. On held-out targets it reproduced those labels 83.2% of the time, ahead of both heuristics. The labels are model judgments rather than chemist preferences, and no deployed Ordo endpoint exists.

Probatio estimates the feasibility of a proposed reaction step, reaching a test AUC of 0.831 with calibration error 0.017 after Platt scaling. Enabled inside the search it reduced top-1 route recovery and nearly doubled median search time, and is disabled in production. A variant combining the fingerprint with the forward model's own confidence signals reached a top-1 AUC of 0.7975, below the 0.816 obtained from the forward model's beam share alone.

Plotted against search time, the four internal configurations and the external baseline separate along the expected trade-off: our configurations achieve higher top-1 recovery at lower latency, while AiZynthFinder performs better at top-10 recovery.

Forward prediction and conditions

Productum I

Productum predicts the major product from reactants and reagents, following the Molecular Transformer^{15[15]Schwaller, P. et al. (2019) Molecular Transformer: a model for uncertainty-calibrated chemical reaction prediction. ACS Central Science 5(9).}. With beam search over 1,000 held-out patent reactions it reached 59.8% top-1 and 75.2% top-5 accuracy. The 67.1% figure sometimes quoted for this model is greedy validation accuracy, not a test result. Published forward models report higher accuracy on different benchmarks and splits, and no public forward model has been evaluated on our split.

Average accuracy on patent reactions obscures failures on elementary chemistry: hydroxide with an alkyl halide yields the ether rather than the alcohol, which is what the patent data predominantly records. We held out forty textbook reactions and scored them by hand. The deployed model answered 29 correctly; a fine-tune on generated textbook chemistry answered 38, with no loss of patent accuracy. The fine-tune has passed its release gate and is scheduled for the next deployment.

Condicio

Condicio predicts the conditions a reaction was run under. The initial sequence-generation approach, which emitted the full agent list as text, was discontinued. The second version predicts five role slots, catalyst, two solvents and two reagents, assigned by rules using a published solvent list^{19[19]ORDerly: datasets and benchmarks for chemical reaction data.}. The third adds mean-pooled features from our own reaction encoders, raising joint accuracy across all five slots from 7.6% to 8.2%.

The limited improvement from substantially richer molecular representations suggests that label quality may be a larger constraint than the representations evaluated here: patent agent lists are unordered, noisy and frequently incomplete. Condicio is not deployed, and temperature prediction is not established. An earlier figure of 73.1% in our training logs scored each slot independently and credited a catch-all answer, and is not reported as a result. Prior work on condition prediction^{18[18]Gao, H. et al. (2018) Using machine learning to predict suitable conditions for organic reactions. ACS Central Science 4(11). doi:10.1021/acscentsci.8b00357.} used different labels and protocols.

Mechanisms

Motus and the rules-first service

Motus predicts electron-pushing arrow sequences, trained on a published set of labeled patent mechanisms^{5[5]Chen, S., Babazade, R., Kim, T., Han, S., Jung, Y. (2024) A large-scale reaction dataset of mechanistic pathways of organic reactions. Scientific Data 11, 863. doi:10.6084/m9.figshare.24797220.} together with mechanisms generated from textbook rules. An execution engine checks each proposed step against structural and valence constraints before it is returned.

The first version scored 91.5% on reactions as written and 4.5% on the same reactions with atom order shuffled, indicating heavy reliance on positional patterns in the training representation. Training on several writings of each reaction removed most of that effect, and the deployed model, version 1.2, scores 81.0% on the shuffled test.

On a separate benchmark of 194 textbook mechanisms, the deployed model alone found 98. Applying the same rules that generate the training data ahead of the model, and invoking the model only where no rule matches, raised this to 171: 145 from rules and 26 from the model. "Exact" on this benchmark is an automated check of mechanism shape rather than agreement with hand-drawn arrows, and radical examples are scored for coverage only.

Six further training runs evaluated improvements to the learned model. Two failed release gates that were subsequently found to be incorrectly specified, and one failure traced to two input-handling bugs rather than the model. With those corrected, the most recent run met every gate condition but regressed on elementary proton-transfer reactions, answering 4 of 12, and was not deployed. A follow-up run addressing this has no gate result yet.

Six examples from the mechanism benchmark, showing the predicted intermediates and whether each prediction came from the rules engine or the learned model.

Reading chemistry

Cernere I

Cernere I recognizes a single molecular structure from an image. It was evaluated against MolScribe^{30[30]Qian, Y. et al. (2023) MolScribe: robust molecular structure recognition with image-to-graph generation. Journal of Chemical Information and Modeling 63(7).} on 1,933 patent drawings from USPTO-30K, with both tools run by us under identical scoring. MolScribe outperformed Cernere across all four subsets, with the largest gap on bridged structures containing stereocenters. Restricting the comparison to molecules absent from MolScribe's training files does not change this.

Removing stereochemistry from the scoring narrows the gap considerably, which locates most of the error in stereochemical recognition. Cernere's confidence estimates do not identify these errors: on drawings containing a stereocenter, readings above the confidence threshold are correct 56.4% of the time. Athanor therefore runs an independent stereochemistry check on every structure it reads.

Five successive runs added synthetic drawing styles, notations and a stereochemistry focus. Accuracy on synthetic validation data stayed close to 88% throughout, while accuracy on real abbreviated drawings moved from 22.8% to 24.4%. Adding real patent drawings to the training set raised it to 68.4%.

Cernere II

Cernere II changes the output representation rather than the model scale. It predicts atom labels with quantized coordinates, then classifies every atom pair as unbonded or as a single, double, triple, aromatic, wedge, hash or wavy bond, and assembles the molecular graph from both. Stereochemistry is then derived from geometry and wedge direction, which is what the drawing encodes. The model has 54.2M parameters and is initialized from our own earlier encoder weights where shapes match; no third-party pretrained model is used.

Training is in progress at 106,954 of 200,000 steps, with 61.4% exact accuracy on its own validation set at 100,000 steps. That validation set is not the held-out benchmark used above, and no benchmark result is available yet.

Scriptum I

Scriptum recognizes handwriting from pen trajectories rather than images, using a recurrent network with CTC decoding after Carbune and colleagues^{31[31]Carbune, V. et al. (2020) Fast multi-language LSTM-based online handwriting recognition. International Journal on Document Analysis and Recognition 23. arXiv:1902.10525.}. A small model reads atom labels and a larger one reads lines of text. The decisive improvement came from preprocessing rather than capacity: the initial model was sensitive to the uniform stroke timing of the synthetic training data, while tablets report points at their own sampling rate. Aligning preprocessing with tablet input raised label accuracy on tablet-style ink from 91.6% to 98.5%.

Through the deployed pipeline, label recognition reaches 98.5% exact and line recognition 58.7%, with text mode accepting 58.0% of lines at 92.0% precision above its threshold. All of these results are on synthetic ink with held-out letter styles; performance on real human handwriting has not been measured. One discrepancy remains unresolved: the first label model scored 96.5% in its training log and 91.8% through the service pipeline on the same checkpoint.

Spectroscopy

Resonantia I, carbon-13

Resonantia I predicts carbon-13 chemical shifts in chloroform using a 2.5M-parameter graph message-passing network, trained on 814,593 experimental spectra^{7[7]Wang, Z., Zhu, X. (2025) NMRexp: an experimental NMR spectral dataset. doi:10.5281/zenodo.17296666.} and evaluated on 31,147 spectra with scaffolds held out of training. Mean absolute error is 0.692 ppm against 2.487 ppm for a HOSE-style baseline^{32[32]Kuhn, S., Schlörer, N. E. (2015) Facilitating quality control for spectra assignments of small organic molecules. Magnetic Resonance in Chemistry 53(8).} we implemented from the same training spectra.

Published spectra list peak positions without atom assignments, so each spectrum is scored after an optimal one-to-one assignment of predicted shifts to observed peaks. This scoring gives any predictor the benefit of the best available assignment and is applied identically to both the model and the baseline. A wider experimental run reached 0.6951 ppm on validation; it is an experimental checkpoint with no test result.

Resonantia-H, protons

Resonantia-H uses the same architecture for proton shifts. Over the 1,022,711 protons explained by both systems, mean absolute error is 0.0808 ppm against 0.1696 ppm for the baseline, paired on the same spectra. Exchangeable protons on oxygen, nitrogen and sulfur reach approximately 0.35 ppm and are reported separately. Both models predict chemical shifts only; coupling constants and multiplicity are not predicted, and neither model distinguishes stereochemically inequivalent protons.

Ranking candidate structures

A derived tool built on Resonantia ranks known candidate structures given a molecular formula and a measured carbon-13 spectrum. Over 5,000 queries it ranked the correct structure first 94.0% of the time. The evaluation measures candidate ranking, with the correct structure guaranteed to be present in the candidate pool. Whether this endpoint is exposed in the app is not established in the reviewed contract.

Molecular properties

Proprietas I

Proprietas I predicts acid dissociation constants for each ionizable site together with an uncertainty estimate, trained on experimental values from ChEMBL^{8[8]Zdrazil, B. et al. (2024) The ChEMBL database in 2023. Nucleic Acids Research 52(D1).}. On held-out molecules it reached 1.434 pKa units against 1.579 for a ridge baseline, and well ahead of both chance controls. It is not deployed because its uncertainty estimates are not usable: intervals labeled 90% contain the measured value 31.5% of the time. On an external set from Wikidata^{37[37]Wikidata.} error rises to 3.024 pKa units.

Planned: Anulus and Nomen

• Anulus. A retrosynthesis policy aimed at ring-forming steps. No trained results exist.
• Nomen. Structure to IUPAC name, verified by round-tripping the name back through a parser. No trained results exist. The name lookup and checking that Athanor does today comes from a third-party parser, not from a trained model.

The textbook set, with every prediction

Forty textbook reactions were held out from the generator that produced the fine-tuning data. Reaction inputs, reference products and the predictions of both forward-model runs are published below. Several incorrect predictions are chemically reasonable products of the same starting materials, which exact-match scoring counts as errors.

Key findings

Single-step accuracy and route quality require separate evaluation. Resolvere Liber improved single-step accuracy over the deployed template policy, from 26.5% to 32.7%, but reduced top-1 route recovery from 25.2% to 21.8% and nearly tripled median search time when integrated into Texere. Probatio showed the same pattern: a test AUC of 0.831 in isolation, and degraded route recovery once enabled inside the search.

Hybrid systems outperformed the learned model alone. On the 194-reaction textbook mechanism benchmark, running textbook rules ahead of the model found 171 mechanisms against 98 for the deployed model by itself, with 145 of those answered by rules.

Training data composition mattered more than capacity or representation. Atom-order augmentation raised mechanism accuracy on shuffled inputs from 4.5% to 81.0%. Adding real patent drawings to structure-recognition training produced a larger gain than five rounds of additional synthetic drawing styles. For handwriting, aligning preprocessing with real tablet input mattered more than model size. For condition prediction, adding encoder features changed joint accuracy by under a point.

Prediction accuracy is not sufficient without calibrated confidence. Cernere I is correct 56.4% of the time on confident readings of structures containing a stereocenter, and Proprietas I achieved 31.5% empirical coverage from nominal 90% intervals. The first result motivated an independent stereochemistry check in the app; the second prevented deployment.

Release criteria did not cover everything that mattered. The most recent mechanism run met every gate condition it was set and then regressed on elementary proton transfers, which no gate tested. It was not deployed.

Training infrastructure

These models were trained on arbastro infrastructure: an NVIDIA RTX 5090, an NVIDIA RTX 5070 Ti, an NVIDIA Tesla T4, and the NVIDIA GTX 1050 Ti in the arbastro Cloud Initium I node in Tampa, Florida. Training runs on a scheduled overnight window, with evaluation on separate machines so that a long training job and a benchmark never compete for the same card.

Each experiment was recorded with its expected outcome and failure criteria before it ran, and outcomes were appended afterward without editing the original prediction. The full record is below.

What is not measured

A retrosynthesis-to-forward round-trip evaluation, in which each planned step is checked by predicting it forwards, has a written protocol and a smoke test but no result. It is listed here as outstanding rather than substituted for.

The following direct comparisons with external tools have not been run:

• Productum against a public forward model on the same split.
• Resolvere and Liber against a public single-step model standalone.
• Condicio against an external conditions model.
• Scriptum against a handwriting recognizer on real human ink.
• Motus against an external mechanism predictor.

A comparison against a commercial NMR predictor was proposed and withdrawn before measurement, and no results exist for it.

Use in Athanor

In Athanor these models are reached through ordinary editor tools: retrosynthesis and product prediction on a drawn reaction, a mechanism tool that checks or proposes arrows, structure import from an image, handwriting input on a tablet, and predicted NMR alongside a structure. Predictions are labeled with their source, and the app validates valence and stereochemistry independently of any model. The registry read on November 12, 2026 reported service version 0.5.0 under contract 0.57.

Not yet on this page: screenshots of these features in the current app.

Limitations and related work

Retrosynthesis results are difficult to compare across publications because protocols, stocks and scoring differ^{24[24]Maziarz, K. et al. (2023) Re-evaluating retrosynthesis algorithms with Syntheseus.}, which is why every comparison here was run by us on shared examples rather than taken from a paper. Systems such as ASKCOS^{25[25]Coley, C. W. et al. (2019) A robotic platform for flow synthesis of organic compounds informed by AI planning. Science 365.} extend further than we have into condition recommendation and execution. Tree search with a neural expansion policy^{20[20]Segler, M. H. S., Preuss, M., Waller, M. P. (2018) Planning chemical syntheses with deep neural networks and symbolic AI. Nature 555.} and template-free reaction prediction^{12[12]Liu, B. et al. (2017) Retrosynthetic reaction prediction using neural sequence-to-sequence models. ACS Central Science 3(10).,15[15]Schwaller, P. et al. (2019) Molecular Transformer: a model for uncertainty-calibrated chemical reaction prediction. ACS Central Science 5(9).} are established methods; what this report adds is their measured behavior inside one integrated application.

Model cards and run history

One card per family member, followed by the full run inventory. Directories holding only data preparation or evaluation artifacts are marked as such rather than counted as models, and the search engine and atom mapper have no parameter counts because neither is a trained network. Mapper reproduces RXNMapper's atom assignments^{16[16]Schwaller, P. et al. (2021) Mapping the space of chemical reactions using attention-based neural networks. Nature Machine Intelligence 3.} on 79.7% of 1,000 held-out reactions. Both systems are compared against each other rather than against independently established ground truth.

Changelog

Model changes, newest first, dated by the result file or recorded decision each entry comes from.

1. Forward prediction passes its textbook gate

   November 12, 2026

   A fine-tune on generated textbook chemistry took the held-out textbook set from 29 of 40 to 38 of 40 without costing patent accuracy, and is recorded as serving with the next deploy. The registry still answered version 1.0 when this page read it.

   Source

   results/models/productum-tb-1/{textbook-heldout,patent-paired}.json

2. Two input bugs found in the mechanism service, and a run that still could not ship

   November 12, 2026

   Lewis-acid catalysts were reaching the mechanism model unnumbered, and some aromatic proposals were opening the ring they attacked. With both fixed, the latest run passed every bar it had been set. It was still not served: it answers only 4 of 12 plain proton transfers, and none at all for phenol with hydroxide. A follow-up trained to restore them has no gate result yet.

   Source

   results/models/mech-9/gate-ringfix.json, results/models/motus-proton-2026-11-12.json

3. Proton NMR served

   November 11, 2026

   The proton shift model was tested once against a paired algorithmic baseline on the same spectra, at 0.0808 ppm against 0.1696, and served. Exchangeable protons are reported separately at about 0.35 ppm.

   Source

   results/models/resonantia-h2/{paired-vs-hose,test}.json

4. Carbon-13 NMR served, and a pKa model held back

   November 10, 2026

   Carbon-13 shift prediction reached 0.692 ppm against an algorithmic baseline at 2.487 ppm on the same data, and was served. The pKa model beat its ridge baseline on the mean and was not served, because intervals labeled 90% covered only 31.5% of measured values.

   Source

   results/models/resonantia-1/test.json, results/models/proprietas-1/test.json

5. Real patent drawings in training, and the image reader at 1.3

   November 9, 2026

   Five runs of added synthetic drawing styles produced little change on real drawings. Training on real patent drawings raised clean-drawing accuracy from 59.0% to 79.0% and abbreviated drawings from 24.2% to 70.8%, across two runs deployed as 1.2 and 1.3.

   Source

   results/models/{imago-real-1,cernere-real-2}/uspto30k.json

6. More synthetic drawing styles stopped paying

   November 5, 2026

   Two further runs, one adding drawing styles and one targeting stereochemistry, changed real-drawing accuracy by amounts within measurement noise. Neither was deployed.

   Source

   results/models/{imago-mixed-5,imago-stereo-1}/uspto30k.json

7. Ink preprocessing matched to real tablet input

   November 3, 2026

   The label reader was sensitive to the uniform stroke timing of synthetic ink. Resampling and re-timing input to match training, then retraining through that path, raised tablet-style accuracy from 91.6% to 98.5%. Deployed as 1.2.

   Source

   results/models/scriptum-label-2/compare.json

8. The feasibility filter stays off

   November 3, 2026

   A step-feasibility classifier with a standalone test AUC of 0.831 was evaluated inside the route search on 500 targets. It reduced top-1 route recovery and nearly doubled median search time, and is disabled.

   Source

   results/texere/n1-p31-{base,probatio}.json

9. Conditions, third attempt

   November 2, 2026

   Adding pooled features from our own reaction encoders moved every condition slot by one or two points, and joint accuracy across all five slots from 7.6% to 8.2%. The limited gain points to label quality as the larger constraint. Not deployed.

   Source

   results/models/condicio-3/test-strict.json

10. Template-free retrosynthesis measured, and not served

    October 29, 2026

    The template-free model outperformed the deployed policy on single steps, 32.7% against 26.5% at top-1. Integrated at the first disconnection of the search it lowered top-1 route recovery from 25.2% to 21.8% and tripled median search time.

    Source

    results/models/liber-1/beam-296000.json, results/texere/n1-p13-liber.json

11. The mechanism model stopped reading atom order

    October 27, 2026

    The first mechanism model scored 91.5% on reactions as written and 4.5% with atom order shuffled. Training on several writings of each reaction raised the shuffled score to 80.0%, and a textbook fine-tune deployed as 1.2 reached 81.0%.

    Source

    results/models/mech-{1,2,3}/order-test.json

12. Forward prediction finished and served

    October 26, 2026

    Forward prediction completed 400,000 training steps. Beam search over 1,000 held-out reactions reached 59.8% top-1 accuracy.

    Source

    results/models/productum-1/{progress,eval-beam-tta}.json

13. Atom mapping of our own

    October 26, 2026

    Template reversal with a maximum-common-substructure fallback reproduces the reference mapper on 79.7% of 1,000 held-out reactions, and on 95.1% of the two thirds handled by the template path. Deployed.

    Source

    results/models/mapper-1/eval.json

14. The route search benchmarked against the open-source planner

    October 23, 2026

    On all 10,000 targets of the public n1 set, with the same stock, our search solved 98.0% against the baseline’s 96.7% and ranked the recorded route first for 27.1% against 20.0%. The baseline recovered it within ten more often, 60.4% against 53.6%.

    Source

    results/texere/n1-r11-full.json

15. The single-step policy, and the ceiling behind it

    October 21, 2026

    Three template libraries were trained and tested on the same 10,000 reactions. Generalizing the templates raised top-50 accuracy from 58.8% to 66.2% and lowered top-1 from 26.0% to 24.7%. The intermediate version is deployed.

    Source

    results/models/resolvere-{1,1.1,1.2}/test.json

Data downloads

The measurement data behind the figures in this report is published as files that can be read without access to our repository. They contain evaluation results and recorded predictions, not model weights.

• Textbook predictions (JSON): the forty held-out reactions, what each run predicted, and a depiction of every structure.
• Training curves (JSON): validation over training steps for every run with a recorded curve.
• Prediction record (JSON): every preregistered prediction, when it was written, and the outcome recorded against it.

Each file records its own sources. The figures above are built from the same values, and a build-time check re-reads the research record and fails if any figure here has drifted from it. Licensing of the underlying datasets remains with their authors, credited in the references.

References

1. 1.Lowe, D. M. (2012) Extraction of chemical structures and reactions from the literature. PhD thesis, University of Cambridge. doi:10.6084/m9.figshare.5104873.link
2. 2.Schneider, N., Stiefl, N., Landrum, G. A. (2016) What's what: the (nearly) definitive guide to reaction role assignment. Journal of Chemical Information and Modeling 56(12).
3. 3.Genheden, S., Bjerrum, E. (2022) PaRoutes: towards a framework for benchmarking retrosynthesis route predictions. Digital Discovery. doi:10.5281/zenodo.6275421.link
4. 4.Kim, S. et al. (2023) PubChem 2023 update. Nucleic Acids Research.
5. 5.Chen, S., Babazade, R., Kim, T., Han, S., Jung, Y. (2024) A large-scale reaction dataset of mechanistic pathways of organic reactions. Scientific Data 11, 863. doi:10.6084/m9.figshare.24797220.link
6. 6.Morin, L. et al. (2023) MolGrapher: graph-based visual recognition of chemical structures.
7. 7.Wang, Z., Zhu, X. (2025) NMRexp: an experimental NMR spectral dataset. doi:10.5281/zenodo.17296666.link
8. 8.Zdrazil, B. et al. (2024) The ChEMBL database in 2023. Nucleic Acids Research 52(D1).
9. 9.Segler, M. H. S., Waller, M. P. (2017) Neural-symbolic machine learning for retrosynthesis and reaction prediction. Chemistry: A European Journal 23(25).
10. 10.Coley, C. W., Rogers, L., Green, W. H., Jensen, K. F. (2017) Computer-assisted retrosynthesis based on molecular similarity. ACS Central Science 3(12).
11. 11.Dai, H., Li, C., Coley, C., Dai, B., Song, L. (2019) Retrosynthesis prediction with conditional graph logic networks. NeurIPS 2019.
12. 12.Liu, B. et al. (2017) Retrosynthetic reaction prediction using neural sequence-to-sequence models. ACS Central Science 3(10).
13. 13.Zhong, Z. et al. (2022) Root-aligned SMILES: a tight representation for chemical reaction prediction. Chemical Science.
14. 14.Tetko, I. V., Karpov, P., Van Deursen, R., Godin, G. (2020) State-of-the-art augmentation methods for molecular prediction. Nature Communications 11.
15. 15.Schwaller, P. et al. (2019) Molecular Transformer: a model for uncertainty-calibrated chemical reaction prediction. ACS Central Science 5(9).
16. 16.Schwaller, P. et al. (2021) Mapping the space of chemical reactions using attention-based neural networks. Nature Machine Intelligence 3.
17. 17.Vaswani, A. et al. (2017) Attention is all you need. NeurIPS 2017.
18. 18.Gao, H. et al. (2018) Using machine learning to predict suitable conditions for organic reactions. ACS Central Science 4(11). doi:10.1021/acscentsci.8b00357.link
19. 19.ORDerly: datasets and benchmarks for chemical reaction data.
20. 20.Segler, M. H. S., Preuss, M., Waller, M. P. (2018) Planning chemical syntheses with deep neural networks and symbolic AI. Nature 555.
21. 21.Chen, B., Li, C., Dai, H., Song, L. (2020) Retro*: learning retrosynthetic planning with neural guided A* search. ICML 2020.
22. 22.Genheden, S. et al. (2020) AiZynthFinder: a fast, robust and flexible open-source software for retrosynthetic planning. Journal of Cheminformatics 12.
23. 23.Yu, Y. et al. (2024) Double-ended synthesis planning with goal-constrained bidirectional search.
24. 24.Maziarz, K. et al. (2023) Re-evaluating retrosynthesis algorithms with Syntheseus.
25. 25.Coley, C. W. et al. (2019) A robotic platform for flow synthesis of organic compounds informed by AI planning. Science 365.
26. 26.Wender, P. A., Verma, V. A., Paxton, T. J., Pillow, T. H. (2008) Function-oriented synthesis, step economy, and drug design. Accounts of Chemical Research 41(1).
27. 27.Burns, N. Z., Baran, P. S., Hoffmann, R. W. (2009) Redox economy in organic synthesis. Angewandte Chemie 48(16).
28. 28.Badowski, T., Molga, K., Grzybowski, B. A. (2019) Selection of cost-effective yet chemically diverse pathways from the networks of computer-generated retrosynthetic plans. Chemical Science 10.
29. 29.Butters, M. et al. (2006) Critical assessment of pharmaceutical processes: a rationale for changing the synthetic route. Chemical Reviews 106(7).
30. 30.Qian, Y. et al. (2023) MolScribe: robust molecular structure recognition with image-to-graph generation. Journal of Chemical Information and Modeling 63(7).
31. 31.Carbune, V. et al. (2020) Fast multi-language LSTM-based online handwriting recognition. International Journal on Document Analysis and Recognition 23. arXiv:1902.10525.link
32. 32.Kuhn, S., Schlörer, N. E. (2015) Facilitating quality control for spectra assignments of small organic molecules. Magnetic Resonance in Chemistry 53(8).
33. 33.RDKit: open-source cheminformatics.link
34. 34.Genheden, S., Bjerrum, E. route-distances.
35. 35.Lowe, D. M. OPSIN: Open Parser for Systematic IUPAC Nomenclature.
36. 36.Paszke, A. et al. (2019) PyTorch: an imperative style, high-performance deep learning library. NeurIPS 2019.
37. 37.Wikidata.link