{"content":"dev journal — fix: rebalance follow-up gate to precision (59c0136f, branch fix/follow-up-precision). Part 2/3 of the 待跟进 optimization. CEREBELLUM_SESSION_FOLLOW_UP_THRESHOLD 0.35 → 0.5: owner decision 2026-10-31 revisits the 2026-10-23 recall bias — the pending bucket was 74% scheduled runs whose bots had already written \"no action needed\", and those runs now carry the explicit [FOLLOW_UP: owner|none] tag the gate consumes before the Noul, so the cookbook 0.5 bar returns; force patterns stay as the recall net. Instructions rewritten to be owner-scoped: a monitoring/inspection report counts ONLY when unresolved failures block on the OWNER (not items another bot/person already owns); explicit carve-outs for third-party waits, conditional offers (\"I can also… if you want\"), and replies that state no human follow-up is needed; closing rule replaced — judge whether the OWNER is asked to act, not whether anything anywhere is unresolved (the old \"when in doubt, prefer follow-up\"). Eval script header + threshold pins updated; new test asserts the 0.35–0.49 band no longer parks sessions. 18/18 gate tests green.","contentType":"text/plain;utf-8","attachments":[],"quotePin":""}