The last request is not an insight

On a first run, REM wrote a project page with three neat Insight bullets. They said what the owner had asked for, what the owner said had been published, and what the owner wanted fixed next. Each had a citation. None answered the question a returning owner would bring to that page: What did this project actually become while I was away?

The writer had been given the owner's messages to coding agents. Those are valuable evidence of intent, decisions and reported results. They are a poor sole source for implementation state. A request to test an init flow is not a test result. A note saying “publish this” is not proof that a package reached PyPI. When a page cannot compare intent with any observed state, it often turns the most recent request into a polished summary and calls it insight.

This preview changes the input before changing the prose. The first-run writer receives a small, local evidence packet from the recorded project: the README, package metadata, checkout freshness, the current local ref and recent commit subjects. Each item can be cited separately. The writer must compare sources when they support a comparison and keep the boundary between a commit and a verified outcome. A stale checkout next to newer commits is a finding; “the work shipped” would still be a guess.

The page also stops reserving space for answers it cannot give. The map may show a full unfinished scaffold, because that communicates what REM has yet to investigate. Once a project is written, an optional section containing only Unknown disappears. The core state, next threads, uncertainty and sources remain. In one private first-run comparison, the written project went from seventeen sections and seven Unknown values to twelve sections and none. Its lead compared the checkout with the current line instead of merely repeating the owner's latest request.

This is a useful page, not a finished memory system. That real page still had no internal links, and the owner's own History still lacked milestones. A good first viewport needs a finding worth noticing; a mature REM also needs the evidence trail and connected pages that let the owner test that finding.