Verifying yesterday's playtest findings against the source before acting showed two were wrong. The nudge is not truncated - it is wrapped to the sidebar before the truncating helper, and an existing test already asserts the sidebar carries 'delegate M1 work'. And 'read' never promised the day job; help enumerates what it covers. Both are struck in place in the report and dropped from the findings queue. The two real changes: a lone 'delegate M1' - the natural near-miss on a two-row instruction - now answers 'needs a mode: work, think, or lie' instead of reporting the player's own host as an unknown mode word; and a starving sink names the rival for its flow, so the line that sat inert for 1200 ticks now reads 'the core is drawing it to research'. Defense: amends agent-play.md's delegate clause and digital-read.md with a binding starving-sentence clause (a rising state names its competitor where one is identifiable, from recorded facts only). Both verified by new tests and observed in a live agent session.