The wrong answer was in your docs the whole time.
Enterprise AI mostly answers by reading your own documentation. When that documentation disagrees with itself, the model cannot tell which version is true, so it guesses, blends, or invents. The industry has a quiet name for this failure: conflict. It is the most fixable reason AI gets things confidently wrong, and almost no one is fixing it.
Pelcrow is how you fix it. It catches the contradictions before your AI ever reads them.
To reset the device, press and hold for 5 seconds.
To power cycle the unit, hold for 10 seconds.
Hold the button for 5 5 to 10 seconds to power cycle the device to reset it. (Which page is current? The model can't know.)
Pelcrow keeps your documentation from arguing with itself.
Pelcrow is the governance layer that sits between your writers, your AI, and everything you publish. It is not the writer. It is the reference desk and fact-checker standing behind the writer: it holds what already exists, which terms are approved, and which version each fact belongs to. Every change, whether a person typed it or a model generated it, passes through a gate that catches contradictions, broken references, and drifted terminology before the AI that answers your customers ever reads them.
What a conflict actually is
Picture a simple question to your support assistant: how do I reset the device? To answer it, the AI does not open one page and read it the way a person would. It pulls every passage in your documentation that looks relevant and reads them all at the same time.
One passage, written last year, says press and hold for five seconds. Another, written after the last hardware revision, says ten. A third calls it a power cycle instead of a reset, so the model is not even certain it is the same action. Three passages. Three answers. All from your own trusted content.
That disagreement is a conflict. The AI has no reliable way to know which passage is current, which is deprecated, or that two of them describe the same thing in different words. So it does one of three things, and all three are bad. It picks one and states it with total confidence. It averages them into an instruction that was never true. Or it fills the gap with something invented. That last one is the hallucination everyone worries about, and contradiction is one of its largest and least discussed causes.
Why you have never noticed
Conflicts have always lived in documentation. They stayed invisible because humans read one page at a time. You land on the current page, you get the right answer, and the stale page three clicks away never enters your field of view. Your reader's attention was quietly doing the filtering for you.
An AI removes that filter. It has no sense of which page you would have landed on. It sees the whole contradictory pile with equal weight and has to resolve it in a fraction of a second. The messiness that was always in your content, harmless for twenty years, becomes the exact thing that makes your AI unreliable the moment you point a model at it.
The part you control
Hallucination is certain. This is not.
Here is the truth every serious AI team accepts: a model will hallucinate. It is a statistical certainty, not a bug you can patch to zero. So the real work is not eliminating hallucination. It is removing every avoidable cause, until what remains is as small as it can be.
Most of the industry spends that effort on the model. Better prompts, better retrieval, bigger context windows, reranking. All useful, and all working on the half of the problem you do not control. The other half, the content the model reads, is the half you own outright. Consistent terminology, one source of truth per fact, and reuse instead of copy and paste do not make your docs prettier. They remove the contradictions that produce confident wrong answers. They are hallucination controls that happen to look like editorial standards.
What changed, and why now
None of this mattered much three years ago, because documentation was slow and expensive to produce, and a person usually owned its coherence. AI changed both facts at once. Writing a page is nearly free now. Anyone, and any tool, can generate one. At the same time, the teams who used to hold the whole body of content in their heads and keep it agreeing with itself got smaller, or disappeared.
So you now have more content, produced faster, by more hands and more models, with less oversight than ever, being read by machines that are ruthless about contradiction. The oldest problem in documentation got worse at the precise moment it started to carry real business risk.
Not the writer. The reference desk and the fact-checker behind it.
Pelcrow does not try to write your documentation. Your own AI can draft. Pelcrow is the thing standing behind it, holding the ground truth: what already exists, which terms are approved, what is reused and where, and which version each fact belongs to.
Every change, whether a person typed it or a model generated it, passes through a gate before it is allowed in. The gate catches the contradictions, the broken references, and the drifted terminology while they are still cheap to fix, long before they reach the AI that answers your customers. What comes out the other side is the one thing enterprise AI actually needs and rarely gets: content that is correct, consistent, reusable, and traceable enough to be trusted.
You will put AI in front of your customers. The only question is which errors you removed first.
It will answer in your company's name and your company's voice. It will sometimes be wrong, because that is the nature of the tool. What you are accountable for is not perfection. It is whether you took away the errors you could have taken away.
Contradictory documentation is the most fixable of those errors, and today it is almost entirely unaddressed. Fix it, and you have removed one of the largest and quietest reasons your AI gets things wrong. That is what Pelcrow is for.