Your AI Didn’t Fail. It Read What You Wrote.
Imagine this. The assistant quoted a client $500. The price has been $750 since March. The client accepted on the spot, the team ate the difference, and the postmortem blamed the AI.
The postmortem charged the wrong defendant. The AI did exactly what it was built to do. It read the folder it was pointed at, found a document titled “Pricing,” and answered with total confidence. The document was last year’s. Nobody marked it dead. Nobody wrote down the decision that killed it. The new price lived in a meeting, a Slack thread, and one updated doc that contradicted the old one without replacing it.
The model didn’t hallucinate. It told the truth about a folder that was lying.
The disease has a name
Call it context rot. Every organization that wires an AI assistant into its documents is accruing it right now, mostly without knowing.
Rot looks like this. Two documents disagree about the same fact, and both look current. A plan gets cancelled in a meeting but never in writing. A price sheet, a policy, a deadline goes stale with no marker and no pointer to its replacement. The index, if there is one, forgot half the files. Decisions happened, but the why lives in nobody’s head longer than a quarter.
Humans survive rot because humans carry context in their skulls. You know the old price sheet is old. You were in the meeting. The folder is wrong but you are right, and you quietly route around the difference every day without noticing you are doing it.
The AI was not in the meeting. The AI has no skull to carry the difference in. It has the folder. If the folder lies, the AI lies, fluently, politely, and at scale.
The part nobody wants to hear
A bigger model will not fix this, and that is the part nobody wants to hear. The instinct is to wait. Next year’s model will be smarter, and smarter will mean safer. But model intelligence is not the failing component. Given two contradictory documents and no marker for which one governs, there is no amount of intelligence that resolves the contradiction correctly. There is only a guess. Smarter models guess more fluently. Fluent guessing is worse, because you stop checking. “Point the AI at our docs” is not a safe sentence yet. Not because the AI is weak, but because almost nobody’s docs deserve the trust. The folder was never built to be load-bearing. Now it is. Nobody inspected the beam.
What current truth actually costs
The fix is not a product first. It is a discipline first. Four rules, none of them clever.
One current truth at a time. For any question the folder answers, exactly one document governs. Everything else is history. History is kept, marked, and pointed at its replacement. The moment two documents can both claim to be current, your AI is flipping a coin you cannot see.
Decisions get records. When truth changes, write down what changed, why, and who decided. Not for bureaucracy. For the day a person or a machine asks “why is it this way” and the answer would otherwise be gone.
Logs are append-only. Fixing history by rewriting it is how folders learn to lie. Add the correction. Keep the mistake. The mistake is data.
Audit after every edit. Broken links, stale markers, files the index forgot, documents that contradict each other. Rot accrues per edit, so the check runs per edit. Discipline that depends on remembering is not discipline. It runs in a script or it does not run.
I did not arrive at these rules from theory. My family runs real business interests out of one governed corpus, with AI agents working inside it daily. My wife runs the LLC, I carry the long-term strategy, and a publishing operation rides alongside. Each rule was paid for the expensive way. A stale figure nearly walked into a negotiation. A decision got re-litigated because the why was never written. A folder sat confident in two directions at once. The same week I built the audit, it caught my own version-number contradiction across three files. The author of the discipline failed the discipline, and that is what makes it real. It does not run on trust, including trust in me.
The trigger event is coming for you
Here is the prediction, and you can hold me to the price of being wrong. Within a year, every team running AI against shared documents will have its $500 moment. An agent quotes a dead price, books against a cancelled plan, files under a revoked policy. The failure will be blamed on the AI. The cause will be the corpus. Most teams will buy a smarter model and have the same accident again, with better grammar.
The teams that get out clean will be the ones that treated their folder like infrastructure, inspected and governed, one truth at a time.
The mechanical half of that inspection is now free. I open-sourced the auditor I run on my own corpus. It catches broken links, index drift, stale markers with no pointer, and encoding rot. Python, zero dependencies, exit codes fit for cron and CI. It will not catch a contradiction between two fluent documents, because no script can. That half needs a reader, either you or your AI wired with the right protocol, and the wiring is the part I sell. The script is at github.com/forgedculture/corpus-keeper. Run it on the folder your AI reads. The findings count will tell you whether your $500 moment is already loaded.
Your AI is only as good as the folder you point it at. The folder is yours, and so is the bill.
Artifacts are cheap, judgement is scarce. Per ignem, veritas.