Agentic reaction classification: 14,073 verifiable rules

Daniel Armstrong, Maarten Dobbelaere, Valentas Olikauskas, Helena Avila, Octavian Susanu, Jérôme Waser and Philippe Schwaller submitted "Agentic generation of verifiable rules for deterministic, self-expanding reaction classification" on 1 Jul 2026. The paper describes a fully automated, multi-agent pipeline of large language models that classifies 665,901 US patent reactions and generates verifiable reaction rules under a testing loop.

What did the authors build?

They built a multi-agent LLM pipeline that writes and verifies deterministic reaction rules against a corpus of 665,901 US patent reactions, producing a self-expanding taxonomy that grew from 68 classes to 14,073 classes. The system generates each rule and then tests it against the full patent corpus in a verification loop; successful rules enter a living database and the taxonomy expands on demand.

The framework centers on agents that classify reactions and author the symbolic rules themselves rather than relying on a fixed, human-curated ruleset. That automation addresses chemistry's long tail by letting the ruleset grow as new chemistries are encountered in the corpus.

How well does it perform?

A lightweight fingerprint classifier trained on the generated rulesset classifies 97.7% of unseen reactions, a coverage the paper says matches a leading proprietary classifier while offering finer chemical resolution. The authors report the taxonomy expansion from a standard 68 classes to 14,073 classes as a concrete outcome, and they applied the pipeline across 665,901 US patent reactions to produce and verify the rules.

The paper contrasts the new ruleset with existing fixed rulesets: automated generation plus verification lets the system extend to chemistry outside its original training distribution. The lightweight fingerprint classifier is the production-facing component that labels held-out reactions; the 97.7% figure quantifies its coverage on unseen examples.

Why it matters

The pipeline removes a major manual bottleneck in computer-assisted synthesis planning: encoding deterministic, interpretable reaction rules by hand. By automatically producing verifiable symbolic rules and testing them against a large patent corpus, the method produces what the authors call a "living reactivity database," enabling a ruleset that grows and adapts instead of remaining static. That matters for teams that need interpretable, deterministic transforms for route planning but require coverage over rare or novel chemistries.

Automation that matches a proprietary classifier on coverage while increasing granularity changes the tradeoff between opaque learned models and brittle, manually curated rulesets. The approach provides a concrete path to use generative LLMs to produce symbolic artifacts that are immediately testable and usable in downstream chemistry tools.

What to watch

Look for whether the verification loop and generated taxonomy are released alongside the paper and whether others reproduce the 97.7% classification coverage on independent corpora. Also watch for adaptation of the approach to non-patent reaction datasets and for comparisons that quantify how the finer-grained classes affect retrosynthesis planning quality and route ranking.

Summary data points: the work was submitted on 1 Jul 2026; it processed 665,901 US patent reactions, expanded a taxonomy from 68 to 14,073 classes, and reports a lightweight classifier that labels 97.7% of unseen reactions. The authors position the result as a living reactivity database and a route to turning generative models into verifiable, self-expanding symbolic systems.