Cost-Bounded Evolving Guardrails for Chemistry Protocol Screening via Event-Pair Evidence Retrieval

Translating a published synthesis paragraph into a machine-actionable protocol is now routine, but judging whether the resulting steps are chemically safe on a given bench is not: frontier models recognise equipment at near-perfect accuracy yet fall to 0.28 accuracy on laboratory safety questions, and whole-text screening of a multi-step paragraph cannot point to which condition–effect pair justifies its verdict. We argue that risk lives at the level of event pairs, not paragraphs, and propose UHG, a cost-bounded evolving guardrail for screening heterogeneous chemistry protocols via event-pair evidence retrieval. UHG compiles text, tabular procedures, and figure captions into one event-pair space, retrieves condition–effect pairs as auditable evidence, distils a teacher’s evidence selection into a lightweight reranker, and evolves per-lab decision boundaries under a hard annotation and inference budget. On 50 catalyst-synthesis protocols with step-level risk labels, UHG reaches unsafe F1 of 0.72 at 0.13 of the cost of whole-text GPT-4o screening, outperforming Claude-3.5-Sonnet whole-text screening

Authors

Publication Details

Journal
ChemRxiv
Published
2026-11-12
DOI
https://doi.org/10.26434/chemrxiv.15012079/v1
Primary Topic
Chemical Safety and Risk Management
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Cost-Bounded Evolving Guardrails for Chemistry Protocol Screening via Event-Pair Evidence Retrieval

Nuoan Cheng, Ziyang Pei, Jiarui Luo, Xinyue Fang et al.
ChemRxiv
Chemical Safety and Risk Management
preprint

Cost-Bounded Evolving Guardrails for Chemistry Protocol Screening via Event-Pair Evidence Retrieval

Nuoan Cheng, Ziyang Pei, Jiarui Luo, Xinyue Fang, Yuchen Bai Kunming
preprint en

Abstract

Translating a published synthesis paragraph into a machine-actionable protocol is now routine, but judging whether the resulting steps are chemically safe on a given bench is not: frontier models recognise equipment at near-perfect accuracy yet fall to 0.28 accuracy on laboratory safety questions, and whole-text screening of a multi-step paragraph cannot point to which condition–effect pair justifies its verdict. We argue that risk lives at the level of event pairs, not paragraphs, and propose UHG, a cost-bounded evolving guardrail for screening heterogeneous chemistry protocols via event-pair evidence retrieval. UHG compiles text, tabular procedures, and figure captions into one event-pair space, retrieves condition–effect pairs as auditable evidence, distils a teacher’s evidence selection into a lightweight reranker, and evolves per-lab decision boundaries under a hard annotation and inference budget. On 50 catalyst-synthesis protocols with step-level risk labels, UHG reaches unsafe F1 of 0.72 at 0.13 of the cost of whole-text GPT-4o screening, outperforming Claude-3.5-Sonnet whole-text screening

ChemRxiv
Chemical Safety and Risk Management
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon