Read a paper for the experiments it lets you skip

A paper is worth more as the list of experiments you no longer have to run than as a verdict on your plan.

One paper, the largest panoramic radiograph reader published so far, trained on images from more than a hundred thousand patients (Nature Biomedical Engineering, 2026). I read it twice. The first time I read it as a threat: does this change my roadmap. It did not, and I filed it. Days later I read it again with a different question: what did these authors pay to find out that I was about to pay for again.

Three things.

They fed the same images to their model at three resolutions and the results moved by under two points. My planned arm that scaled the whole image was struck the same afternoon.

They had already measured the arm I was keeping. It stayed, as measured, not re-run.

They outlined every tooth on every image, healthy teeth included, before asking the model about disease, and that lifted diseased-tooth identification by more than a fifth. Where to look was worth more than more pixels. That one went into my plan as a step, not a citation.

The harvest went into the training recipe, at the line where each decision sits, the same day. Not into a note, not into a chat. A harvest written anywhere else is gone in a week, and the next version of the plan will not find it.

Read a paper twice. Once for the verdict, which is usually "unchanged". Once for the harvest. Every experiment in it is one you can skip, and the ablation table, the page where the authors vary one thing at a time, is the most expensive page in the paper. It is also the one most people skim past on the way to the headline number.

What I would do differently: read for harvest first. The verdict came first and the harvest waited days it did not need to.

The check

The harvest pass, in three moves:

1. Find the ablation table, the experiments that vary one thing and hold the rest.
2. For each row, ask: was I about to run this. If yes, strike your arm or keep it with their number, and write down which.
3. Put the change into the recipe or protocol at the line it affects, with the citation, the same day. If it is not at the point of action, it did not happen.

Sources

• Nature Biomedical Engineering, 2026, DOI 10.1038/s41551-026-01713-8.