What Can the Records of Published Self-Driving Laboratories Predict About a New Platform? A Step-Level Database and Explorer

Published closed-loop self-driving laboratories (SDLs) are described in prose, so what their records could predict about a new platform has not been tested. I release sdl-workflows , a database of 237 SDL papers coded by a language model as ordered steps, each with an operation and the type of module performing it, and as connections between modules, each with a transfer mode such as fixed tubing, a robot arm or a person (2610 steps and 1681 connections). On 30 held-out papers, the records agreed with an independent coding by a second model (micro-F1 0.891 to 0.905). The release includes the timing data the papers state and an explorer for comparing platforms. A model fitted to papers up to 2023 assigns each connection the most frequent mode among earlier connections between the same operations in the same field. It predicted the recorded transfer mode of 424 connections in papers from 2024 better than the most frequent mode per pair of operation families (accuracy 0.512 against 0.441; 95% confidence interval of the difference 0.016 to 0.125). A model of the same kind did not predict the module type of 1189 steps better than the most frequent module type per operation. A discrete-event simulation predicted reported throughput no better than the median throughput reported by other papers in the same field. Its inputs are rarely stated: 23 of the 237 papers give the duration of every step. The prediction of module types and transfer modes is released as a benchmark.