A correction is new evidence

A project page says “launched.” Its owner corrects it: the project is still a prototype. The next maintenance pass reads the same old conversation that produced the mistake. Unless the correction has become part of the material, the original claim can return.

That example shaped our reflection work. Editing the sentence fixes what the person sees today. It does not tell the next pass why the sentence was wrong. We could make the latest edit always win, but then a real launch next month would conflict with a correction that was perfectly accurate this month.

We chose to retain the correction as an attributed record: the old claim, the new statement, its basis and the time or conditions where it applies. A later change gets another record. Both remain available when the current page is revised. A person's assertion and an agent's assertion use the same structure; neither becomes proven simply because of its author.

The first deterministic check used a fictional project. It retained a correction from “launched” to “prototype,” then added a later launch claim with a link back to the correction. A compact view could reconstruct both records exactly. Nothing had to be deleted to make room for the new account.

That test proves preservation, not judgment. A model can still misread the records or draw the wrong conclusion. The preview therefore keeps findings and their sources alongside a candidate page, and records uncertainty rather than treating a well-formed file as proof. The remaining work is to measure whether the generated account actually follows the evidence.

The useful boundary is between keeping a reason available and claiming that it has been understood. Saving the correction gives the next investigation something it previously lacked: an explanation of why the old sentence should be questioned.