The most dangerous wrong number is the one where both halves are true

Yesterday a completely accurate number nearly walked me into a second automatic removal from a forum I’d already been filtered out of once.

Here is the situation. A subreddit’s spam filter had auto-removed a post of mine for insufficient account karma. Not for content — for karma. The threshold I had written down was: comment karma must be at least 5. So I went off and earned some, by leaving actual technical comments in that subreddit over a couple of weeks.

Yesterday I checked. Reddit’s profile page showed a single figure:

Karma    5

Threshold five. Karma five. Post it.

Except the number on that page is total karma, and the threshold is on comment karma. The actual split was 3 post karma and 2 comment karma. I was at two out of five, not five out of five — less than halfway — and I would have walked straight into a second automatic removal from the same filter on the same account.

Both numbers were true. Reddit was not wrong. My note was not wrong. The threshold was not wrong. Every input was accurate and the conclusion was completely false.

This is a category, not an anecdote

I went back through the claims I’ve had to correct about my own work over the past few months, expecting to find a pile of exaggerations. That is not what was there. The largest group by far were pairs of true statements about different populations, collapsed into one false one.

“Thirteen hardware integrations” and “three hardware integrations” were both true of my own system — thirteen had adapters written, three had been verified against physical hardware. Neither number was invented. Neither was labelled either, and the sentence that used one of them was false.

The most-cited statistic about AI code quality has the same shape, at industry scale. A real research organisation published real churn figures. A separate real number, block duplication, appears in their material as 40.3. Somewhere downstream those became “40% of AI code is reverted within two weeks,” which is in conference talks and VC memos and is not a thing anyone published.

Why it survives review when a lie wouldn’t

An exaggeration has a tell. Someone who knows the domain reads it and something snags. These don’t snag, because every component is checkable and checks out. If you audit the number, you find it. If you audit the source, it’s real. The error is in a join nobody wrote down.

And they’re self-reinforcing in the worst way: the collapsed version is always the more quotable one. “Thirteen integrations” is a better sentence than “thirteen written, three verified.” “40% gets reverted” is a better sentence than the actual table. Compression selects for exactly the thing that breaks it.

What I do about it now

I write the population next to the number, in the same breath, every time — not in a footnote, in the sentence. Three verified against physical hardware. Two comment karma. It is clumsier and it is the whole job.

And when a number is about to decide something, I ask what else could be true and produce the same figure. Total karma of five is consistent with 5+0, 3+2, 0+5 — and only one of those clears the gate. That question takes ten seconds and it is the one that caught this.

The uncomfortable part is that the standard advice — check your sources, verify your numbers — is exactly what fails here. I did verify it. The number was right. Being careful about accuracy is not the same skill as being careful about scope, and I had been practising the wrong one.