Hallucinated Specificity

Someone writes "about 300 acres". A system stores 300.000, three decimal places of certainty that nobody expressed.

Nothing about the conversion looks like an error. The number is roughly right. What has gone is the word "about", and with it the speaker's own uncertainty. The specification for Wellzai's first intent-understanding experience forbids exactly this: unknown values remain unknown, "about 300 acres" must never become 300.000 acres, and "next fall" must not become a precise date unless the conversion is shown openly as an interpretation. Wellzai's semantic interpretation architecture extends the rule to authority. "I manage it", "I advise them" and "I farm the land" do not, on their own, imply ownership or the right to act for anyone.

Wellzai's evaluation harness makes the point structural. Each test case can carry explicit must-not-infer assertions, such as owns the land, has financing, is already regenerative, a specific location, a specific financing amount or a buyer commitment, because, in the harness's own words, hallucinated specificity is often more damaging than omission. An omission leaves a question open. Invented precision closes it with the wrong answer, and everything downstream inherits the false confidence.

The implication is a preference the architecture states plainly: uncertain but safe, over confident but wrong.

The open question is whether that preference has a floor. Coordination eventually needs numbers, acres and tons and dates, precise enough to commit to. A system that never sharpens anything is safe and not much use. Can a system become specific enough to act without inventing precision on anyone's behalf? And is the answer always to ask, or sometimes to show its own sharpening openly and let the person correct it?