Okay, “lied” is dramatic. It just quietly told me something that wasn’t true anymore, and I almost didn’t notice.

Here’s the setup. Back in 2022, as an undergrad, I did a project on waste management in Atonsu, a neighborhood in Kumasi. Very analog, paper questionnaires, door to door, asking people how they get rid of their trash and whether they’d been sick lately. Three years later I decided to drag it out of the drawer, add real machine learning on top of it, and actually publish it. I liked the idea of returning to old fieldwork with new tools and seeing what it had to say.

I built a Random Forest to predict illness from disposal habits, got some cross-validation scores, wrote a little summary of them in a markdown cell like a responsible person, and moved on to the next thing. Weeks later, I’m writing the actual paper, and I go to copy those numbers into the Methods section. Standard stuff. Except when I re-ran the notebook to double check, because something made me pause, I honestly don’t even remember what, the numbers that came out did not match the numbers I’d written down.

Not close. Different model, basically. At some point between writing that summary and finishing the analysis, I’d tweaked a parameter, turned off a class-weighting setting, and just… never went back and updated the sentence describing it. My notebook was describing a version of the model that no longer existed.

The unsettling part wasn’t the mistake itself. Mistakes happen. It was realizing that if I hadn’t had that nagging feeling to double-check, this wrong number would have sailed straight into a public paper looking exactly as confident and finished as a correct one would have. Nothing about it looked off. That’s the trick with this kind of error, it doesn’t announce itself. It just sits there, indistinguishable from the truth, waiting for someone to actually go check.

So I did the least glamorous thing possible: reran everything, traced every number in every results section back to the exact code that produced it, and found one more thing while I was at it, a sentence about which waste categories my image classifier tended to mix up that wasn’t quite what the actual confusion matrix showed. Fixed that too.

Now I have this slightly paranoid habit where, before I write any sentence with a number in it, I ask “did I just look at this, right now, or am I remembering it from three weeks ago?” It’s a small thing. But I’ve started to think good research is mostly made of small things like this, not big leaps, just refusing to trust your own memory of your own work.

Anyway. The paper’s on arXiv now, numbers verified, no ghosts left in it. But honestly the best part of the whole process was just the specific feeling of catching my past self being sloppy and getting to fix it before anyone else saw.