Where do the confident wrong answers come from?
You stood up a local model. You pointed it at the company folder. It answers in a calm, complete voice — and the number, the step, or the warning is wrong. Most people call that a hallucination and blame the model. The model is usually doing what it was built to do. The folder is feeding it.
The model will always sound sure
It is a completion machine. Ask it a question and it will finish a fluent answer. It draws on two piles: whatever you put in front of it, and everything it was already trained on. When the first pile is empty or rotten, the second pile fills the hole — a typical torque spec, a typical clause — in the same voice as a real citation. That is why the wrong answer feels authoritative. It is supposed to.
Three things in the folder do the damage
Two are rotten source. One is wrong search. Same calm wrong answer either way.
1 · Ghost data
A page that looks ingested and is empty. A scan that is only a picture. A figure with no words. The system counts “400 pages.” The model has a blank and invents a typical number. It looks like content. It isn’t.
2 · Hallucination fuel
A page that came out as broken text that still looks like English or numbers. Doubled letters. Two layers of PDF text on top of each other. OCR junk with a “47” sitting in it. The model quotes the junk as if it were the manual. Corrupt text that invites a made-up number.
3 · The wrong neighbor
Different kind of failure: the source may be fine, but the wrong chunk wins the search. The 2014 procedure answers a 2024 question. A running footer or a wiki banner ranks higher than the real section. The model is still sure. The citation is just the wrong document.
Why an honest leftover list matters
If you hide the holes, the model still fills them and nobody can check. If you leave the junk in, it quotes the junk. The useful move is both: take the junk out of what the model is allowed to quote, and leave a list of what was empty or untrustworthy. A person can read the list. The model can be told not to treat those stretches as fact.
We call that stance residual-honest. Not a slogan. It is the paper trail for a wrong answer. The package we mill is a ZIP you keep — a knowledge brick — with that list written down (LIBRARY_CARD and muted chunk flags: that is the residual board). Not a chatbot and not a rewrite that looks clean.
What teams already tried
Blame the model and buy a bigger one. Add prompt rules. Run an extractor so “we have markdown.” Throw the share at RAG again. Those moves can help at the margins. They do not name ghost data, hallucination fuel, or the wrong neighbor — and they do not put corpus fitness on the cost slide. The confident wrong answer keeps coming from the folder.
The fix is to manufacture that leave-behind — junk off the answer path, listed on a board. Others can do serious work here too; we do it as bin in, ZIP out, or no charge. Drop the pile on the mill when you want that done simply.