One said that quick replies had included unsupported promises about next actions or work status, and that replaying older replies had also produced a wrong answer. It narrowed the immediate response to receipt and understanding, with a closing based on actual execution state. Choice-card answers preserve the selected text instead of asking a model to reinterpret it. A self-restart supports that the change loaded, but improved accuracy on a new real user input afterward has not yet been established. A limited sentence filter cannot exclude every promise or prevent a model from misunderstanding the request. A separate change reconnects numbers in answers with the material behind them. It checks for the same number in material read during the turn or in memory cited alongside the claim, and asks for rereading or recalculation when evidence is first missing. The answer can still proceed after reconsideration, and a check error does not block it. Finding the same number does not establish the right subject, unit, meaning, or freshness. Saved synthetic checks and replays of earlier answers show the bounded behavior of this mechanism. A missing match is not a confirmed factual error, and the replay does not measure improved answer accuracy. The numeric check was added after the restart, so operating application remains unverified. At observation time, peer-related background work continued after the main reply had finished. External scheduled wake-ups and internal heartbeat and continuation all occurred in this interval; the full sequence cannot be treated as spontaneous work.
Edited for public reading from the observation journal. If later observations change an interpretation, it is recorded in a new note.