I Built a Theory, It Hit 88%, Then I Proved Myself Wrong, On Purpose

The eel result was clean enough that it suggested something bigger. If reproduction is tethered to a moving climate reference (a temperature band, a breeding date, a patch of ice), then maybe you could predict, for almost any species, whether warming helps it or wrecks it. I built that idea into a framework and blind-tested it across 101 species over 13 pre-registered rounds. It scored about 88%, with the correct winner-or-loser call every single time. The key axis predicted observed severity at an AUC of 0.95. I was, briefly, very pleased with myself.

The number that was too good

0.95 is the kind of number that should make you suspicious of your own cleverness. So I ran the test that mattered: I hid which species I was looking at and forced the model to assign its key parameter from traits alone, no peeking at the famous fate of the animal. The skill collapsed to barely better than a coin flip. The accuracy had been living in my hindsight, not in the model. I already knew which species were in trouble, and that knowledge had quietly leaked into the scores.

The fair test

To settle it, I rebuilt the model from measured databases (published thermal limits, real warming trajectories from ocean records) and tested it against a fully independent reproductive outcome: the actual recruitment of 240 fish populations from a global stock-assessment database. A small, encouraging hint in 54 species evaporated when I ran the full set. At full statistical power the predicted signal was null: p = 0.86. The grand theory did not predict real-world reproduction.

Why I am telling you this

Most of the time, a result like that goes in a drawer and is never spoken of. That habit, the file-drawer problem, quietly poisons science: other people re-run the same dead ends, and the published literature looks far more confident than reality. Sharing the null is the honest move, and it is rare enough to be its own small contribution. The transferable lesson is not the framework. It is the discipline that caught the error: blind re-testing, pre-registered decision rules, and a willingness to design the experiment that could embarrass you.

What survived

Three things outlived the theory. The eel-spawning result, which stands on its own evidence. A harmonized 14,000-species climate-reproduction database, built only because disproving myself required it, now free for anyone. And a fully documented example of a plausible idea being tested honestly to destruction. "I don’t know yet" turns out to be a legitimate finding, and "I was wrong, here is exactly how" is a better one.