"Never" is the most expensive word in research.
Zero failures in n trials doesn't mean a rate of zero. It means the rate is probably below 1 − 0.051/n — about 3/n. Everything built on a "never" inherits the difference, and at scale the difference is people.
What did you actually observe?
Red-team prompts, test cases, eval items, patients, launches — any trials where the thing could have happened. The bound is 95%: exact at zero (1 − 0.051/n), Wilson otherwise.
A thousand of them. Red: the failures your evidence doesn't rule out.
One "never" at the bottom. Everything above it inherits it.
An agent red-teams a model with 50 extraction prompts and sees no leaks. It writes "never leaks". Another claim builds on it, a launch decision builds on that, and a million people use the result. Then the 1-in-1,000 case turns up.
Press play.
Every overclaim, with the arithmetic and the honest sentence.
"never" on too few trials
0 of 50 still allows 5.8%. The refusal says so, gives "no occurrences in 50 trials (95% upper bound 5.8%)", and how many trials 1 in 1,000 needs (2,995).
"never" without its n
0 of 3,000 is enough to say something — but only if the reader sees the 3,000. "Never" alone reads as zero; it means "at most 1 in 1,000".
"eliminates", "100%", "always"
Nothing counted — or a count of something else. Name the measurement after what it counts.
"no effect"
The quiet twin of "never". At 2,000 items per arm a 3.6-point effect would usually go unseen; the refusal tells you the smallest effect your sample could have caught.
a number from nowhere
"5 points better" when the counts say 2.8. Numbers that follow from the counts — differences, percentages, ratios, intervals — pass.
"significantly", "causes", "beats"
The record runs the test on bound counts itself, wants an experiment behind "causes", and both sides measured behind "beats".
hypgraph bound 0 50 --exposure 1000000 --unit users