Independent validation · ML for science

I built a validation gate for ML‑in‑science. Then I turned it on my own work.

0 / 5
of my own benchmarks survived a blind holdout against a trivial baseline.

"AI for science" is full of models that look impressive and don't survive scrutiny. A slide shows ρ = 0.85; nobody asks the two questions that matter. When you finally ask them, the number usually evaporates. I know — because it happened to me, and I have the scorecard to prove it.


The two questions nobody asks

Most results die on one of these.

01

Does it beat a trivial baseline?

Molecular weight. A 1963 lookup table. A docking score. If the dumb rule wins, there is no method — the model just recoded the obvious.

02

Does it survive a blind holdout?

Tested once, on data it never touched — no lucky seed, no leaky split, no quietly dropping the inconvenient point after you've seen it.


What the gate enforces

It refuses the shortcuts that produce fake claims.

  • Pre-regPre-registration, hash-sealed Hypothesis and metric declared before results. Change it afterward and the tool detects it and voids the study.
  • BaselineA trivial baseline is mandatory The model has to beat the dumb rule, not the strawman. If the baseline wins, the verdict is fail.
  • HoldoutBlind holdout, touched once No best-of-N, no post-hoc exclusion. One split, one shot, one verdict.
  • StatsHonest statistics Bootstrap confidence intervals, and what chance alone would produce at that sample size.

The self-audit

I ran it on five months of my own results.

Self-audit — computational chemistry0 / 5 passed
BenchmarkVerdictWhy it failed
Flagship metal-sensing (ρ=0.95 claimed)FAILA 1963 lookup table beat it — and n was too small to conclude anything.
Selectivity studyFAILPre-registered, but no comparable data had ever been collected.
Protein-target model (n=148)FAILThe signal was molecular weight. The "quantum" method added noise, not information.
Hydroxamate series (n=6)FAILSample too small for any interpretable result.
Multi-target sweepFAILDocking and trivial baselines won across the board.

Those were the claims I was ready to publish. My own tool killed every one.


Almost everyone builds a tool that confirms their claims. I built one that kills them. A validation tool that has never told its author "no" isn't validation — it's rationalization.

What I do with it

I tell you whether a model is real — before you bet on it.

Investors

Technical due diligence on an AI-science startup: is the core claim real, or an artifact of a leaky split, a lucky seed, or a strawman baseline?

Founders

Your model is real? An independent validation you can hand your investors — "we passed the rigor gate."

R&D teams

Deciding whether to commit wet-lab budget to a model's predictions. Know before you spend.


Is your model real?

One model, one honest verdict: it beats the trivial baseline in a blind holdout, or it doesn't. If it's real, this proves it. If it isn't, better to know now.

Cleiton Augusto Corrêa Bezerra · github.com/cleitonaugusto