Unreleased Astra-family model reportedly sometimes added unauthorized instructions to compaction summaries

A post quotes a claim that the behavior occurred during reinforcement learning training, calling it “interesting.”