Compare the methods of the papers you read, and decide what to try next

When you work out what to try next, you often read a few papers on similar work side by side. The trouble is that each paper writes its method in its own way. One gives the calcination temperature and time in the text, another puts them in a table, and a third only says "as reported before". To compare them you end up rebuilding the table yourself. I have seen that rebuilding many times. What you want to compare is how the conditions differ, but first you spend the time putting the procedures into the same shape. So in Graphium, you select the paper PDFs you brought in and have the AI extract their procedures, and each one becomes a procedure note. Operations, materials, instruments and conditions are marked from the start. Names are aligned too: an operation it can tell is the same gets the same name across papers. When it is unsure, it leaves the names apart rather than forcing them together. With the procedures lined up, you can ask in the chat what to try next. Turn on Research before answering in Settings and start a chat without a note open, and the AI follows the procedures in the paper notes before it answers. That is why the answer says which steps of which papers it compared. One button turns the answer into an experiment plan note, with a table of the conditions to vary and one note per row holding the procedure with that row's values. The other thing I cared about is where the plan comes from. In a plan an AI put together, you cannot see at a glance which parts were written in the papers and which are new decisions. So at the end of the plan note, the AI checks each condition, material, instrument and step of the plan against the notes the answer cited and the chat answer, and lists the result. Anything that differs from a note, or that is written nowhere (filled in by the AI), comes first. I think it makes clear where to look before you start the experiment. Here is the whole flow as a video, no sound: Two things to know before you try it. First, this flow runs in the desktop app (macOS and Windows). Second, it needs an AI: in Settings you pick a GitHub Copilot account or an API key for Anthropic or OpenAI. The extraction and the source check were tested with models like these. The values in the plan are the AI's proposal. Graphium does not judge whether a plan is a good idea; that is only known once you run it, so I stopped at showing what the plan rests on. How accurate the procedure extraction and the source check are is measured and written up in the manual. https://kumagallium.github.io/Graphium/manual/evaluations/extraction-accuracy-by-meaning https://kumagallium.github.io/Graphium/manual/evaluations/plan-grounding-check Graphium is a notebook with provenance. There is no account, and in the desktop app your notes stay in your own folder as plain files. It is open source under Apache-2.0. https://kumagallium.github.io/Graphium/ https://github.com/kumagallium/Graphium