Passage A:

A patent must enable skilled readers to make and use the claimed invention without undue experimentation. In chemistry, courts have often tested enablement by asking how many compounds a claim covers and how many examples the patent actually teaches. That arithmetic becomes misleading when machine-learning systems nominate enormous families of molecules. A claim may list millions of candidates yet provide a model that ranks them so accurately that a chemist needs to synthesize only a handful. Conversely, a narrow list may be practically useless if the patent offers no reliable way to identify the member possessing the promised property.

Enablement should therefore focus on search burden rather than claim size. Courts should examine the model's validated error rate, the accessibility of its inputs, the cost of testing top-ranked candidates, and whether the patent discloses enough of the model for others to reproduce its guidance. The legal question remains traditional: what work is left to the skilled reader? New tools change the answer, not the standard.

Passage B:

Search burden matters, but Passage A lets prediction substitute too readily for teaching an invention. A model can rank molecules successfully on the training distribution yet fail in an unexplored chemical region. If a patentee claims that region before making representative compounds, later researchers bear the risk of discovering that the map stops at the border. A low average error rate does not disclose where errors cluster.

Representative examples serve more than a navigational function. They demonstrate that the inventor possessed a principle extending across the claimed territory and give readers anchors for diagnosing failure. Model disclosure cannot always perform that work, especially when training data omit unsuccessful experiments or when small laboratory choices affect results. Courts should consider computational guidance, but broad chemical claims should still require examples distributed across materially different parts of the claimed family. Otherwise, a prediction becomes a reservation of research space rather than an enabled invention.

The disagreement has a temporal dimension as well. A model may improve after filing as more laboratories contribute results, but enablement is ordinarily judged from what the patent taught at filing. Later success can confirm that a disclosed principle worked; it cannot supply regional guidance that the document withheld. Nor should a single failed example defeat a genuinely robust method. The difficult cases require courts to distinguish routine debugging, which skilled readers may perform, from a research program that effectively asks the public to discover which portions of the claim the inventor actually possessed.

Disclosure of negative results is especially probative because it reveals the model's boundary conditions. Passage A would treat such disclosure as improving reproducible guidance. Passage B would still ask whether actual examples occupy the regions where the disclosed failures make prediction least secure.