What research taught me that papers didn't

My first week at ICAI Research I implemented a U-Net architecture from scratch, following a paper. In the paper it worked perfectly. On my data, the model converged to zero-loss without learning anything useful. The difference: the paper didn't mention that their data had been preprocessed for six months.

This taught me something I still apply: papers are best results. Authors publish when something works, in the exact configuration in which it works. Everything that went wrong during months of iteration is compressed into "we found that X was crucial."

I'm not saying papers aren't useful. They're essential. But I learned to read them as maps of terrain that someone covered before me, not as GPS instructions. The terrain may have changed. My data isn't theirs. Neither is my hardware.

What was hardest to learn was failing with rigor. A well-documented failed experiment is worth more than ten successes with no record. When something doesn't work, the useful question isn't "why doesn't it work?" but "what is this telling me about the problem?"

That's what I try to preserve when I document code now: not just the what, but the why of things that don't seem to make sense at first glance. Someone — me in six months — is going to need it.