A folder is a route, not a proof

The first Project memories in a private trial looked useful. They described what the user wanted to collect, what they had asked to filter, and where a later folder needed organizing. A source audit caught two opposite ways a page could mislead. One candidate tied old work to a newly linked subfolder without a message making that connection. Another audit could see each user message but not the session folder that had led the mapper to place it on this Project page. It treated the whole older conversation as unrelated.

The distinction matters. A coding session's mapped folder explains why its inputs appear together. It does not prove the requested system exists, that someone ran a crawler, or that a later subfolder contains the older work. We now give the audit that folder provenance alongside each cited input and show it in the owner's source dialog with the same limit. The page can describe folder-scoped conversation history while leaving implementation and subfolder identity unknown.

We also stopped using citations as a mere formatting check. The Project candidate must cite an openable original, and a separate audit compares each claim with its adjacent source. A request for a list is still a request; permission to commit does not show what was committed. When one audit finds a problem, the writer may revise the candidate a bounded number of times. A failed candidate stays available for inspection while the previous page remains intact.

This approach costs extra model turns and can still refuse a useful page. It buys a clearer boundary between what the notebook observed and what it would be tempting to infer. The next test is a broader set of written Project pages with source-by-source review at desktop and phone widths. A single passing page cannot establish whole-notebook accuracy.

The 1.9.0a31 preview notes explain the shipped guard and the single-page retry command.