Episode 4 of Decide with Jev is out!
A recommendation only helps if you know which draft it's about. Say a check comes back "ready for editor". Then someone edits the draft. The old label is still on the screen, and it still looks current.
In this episode, we connect the checks to a small review screen. Every check saves a run with the text we sent and the answers we got back. The page also keeps a hash of the brief and the draft. When the text changes, the old result shows up as stale until you run the check again.
The editor's choice gets its own field. It sits next to the model's answer and doesn't replace it. If the editor disagrees, both stay on the page, along with the reason.
The video also shows:
- why every save gets a new version, even when the text is the same
- how a hidden run ID lets the server reject an old form from another tab
- a session lock, so a save and a check can't write over each other
- a quick test in the browser with a complete draft and a thin one
It ends on one gap we still have to close. That's where episode 5 picks up.
The code is on the episode-4 branch of the Decide with Jev repository. You can watch the whole series in the playlist.