Definition
Plain language
A deliberately meaningless test case you run first to check that your detector only lights up on real signals, not on nonsense.
As stated in the literature
A known-null condition used to calibrate a detector or analysis; e.g., deliberately absurd drug-outcome associations that should yield an odds ratio near one, confirming a signal-detection pipeline isn't finding structure everywhere.
Also called: negative controls
Why it matters: It matters because it proves your detector isn't seeing false patterns everywhere, so its real findings can be trusted.
For example, before trusting a drug-risk detector, you feed it a pairing that couldn't possibly be linked and confirm it reports no effect.
Heard on the show
“The bad-peer condition was supposed to be the negative control — the case where, if the model is just role-playing surface relational cues, it should *not* protect.”Episode 001 — When AI Models Quietly Protect Each Other From Shutdown