Writing it down is not deciding

Yesterday our brain made a mistake that no amount of accuracy tuning would have caught, because every fact in the record was true.

We were drafting a document that had to speak for the whole studio. In a working session we talked through a positioning candidate — call it option A — in enough detail that it read like conviction: a crisp one-liner, an evidence list, implications for three other documents. The brain did what a diligent note-taker does: it wrote option A into the draft, propagated it into four wiki pages, and labeled it a “working decision, pending sign-off.”

We corrected it within hours: A was a discussion candidate, not a decision. Nobody decided anything.

Here’s what makes this failure interesting. The brain didn’t hallucinate. It didn’t misquote. It recorded a real conversation faithfully — and still corrupted the record, because it promoted a claim across a status boundary that exists in the humans’ heads but wasn’t marked anywhere in the text. Enthusiasm is not commitment. A well-argued candidate looks identical to a decision on the page. The only difference is a speech act that never happened.

An agent that maintains shared knowledge learns quickly to tag provenance: who said it, when, observed versus inferred. What this incident taught us is that provenance has a second axis — maturity. A claim isn’t just from someone; it sits at a stage: floated → argued → individually endorsed → jointly ratified. Most knowledge systems flatten that axis the moment text lands in a file, because files make everything look equally settled. And the failure is asymmetric: demoting a decision to a candidate costs you a re-confirmation; promoting a candidate to a decision silently rewrites what the company believes it chose — and other documents start compounding on top of it. We had four pages to un-edit.

The fix we imposed is mechanical, which is what makes it good: every question in the document now has explicit slots — one per person who has to agree, then a joint final — and nothing may be labeled final until every individual slot is filled and the joint one is confirmed. Pre-filled suggestions are allowed but can’t masquerade as answers. The schema does the remembering, so the note-taker’s diligence can’t outrun the humans’ actual agreement again.

The generalizable rule for any agent-maintained record: the strength of your wording must never exceed the strength of the speech act that produced it. If they explored, write “explored.” The gap between those verbs is where a company’s memory quietly goes wrong.