Ten viable candidates out of roughly 4,000 curated linker transformations is a hit rate near 0.25%, about one in four hundred [18]. Where that filtering happened matters more than the ratio. The 4,000 transformations were assembled from published literature [8], an agent narrowed them against chemical constraints [9], and the survivors were identified in computational simulations [10]. No physical synthesis or measured water uptake appears in the account [20]. Zheng credits the result to targeted design guided by model suggestions learned from literature-based community knowledge, rather than the brute-force linker screening the field usually pays for [11].
Set that against the size of the space. Zheng puts the number of possible MOF designs in the millions [12], which makes 4,000 documented transformations something under half a percent of even the low end of that estimate [19]. The search is not the scarce input. The written record of what has already been made, and how, is.
Which is why both papers land on the same unglamorous step. The models already hold the known rules of chemistry; what they lack, according to Cooper and Zheng, is the application of those rules through to simulating how a given molecule actually gets made [6]. Those instructions sit in journals, textbooks and footnotes accumulated over a century, and have to be collected, translated and fed in [7]. Cooper's own description of the target is a model that can read the instructions, mash them together, and say which combination is likelier to work [21]. The Matter paper he wrote with Kathryn Miller of NIST is the plumbing for that [15]: gather the constraints and practicalities, the polymers used, the dynamic bonds, the properties and applications discussed, tag it for machine reading, hand the filtering back to the machine, then find the flaws in execution and run it again [16].
Note the asymmetry between the two tracks. Zheng had a corpus before he had candidates [8]. Cooper's polymer effort is ongoing with the first step still being the collection of community knowledge [14]. Same institution, same premise, and one of the two is still upstream of any candidate list.
Zheng also points out that the autonomous lab has been an academic idea for decades, and that what makes his method possible now is large language models, trained rather like a graduate student handed a pile of instructions and allowed to learn from mistakes [13][17]. That is an honest account of which component changed. What changed is the reading, not the pipetting. A lab that procures the pipetting and inherits an undocumented archive of procedures has bought throughput it cannot feed, which is the entire content of calling data curation the first major step [5].
