Negative exchanges have consequences: the character replies later, then in fewer words, then not at all. We describe the thresholds, the recovery path, and why an always-available model reads as a service rather than an acquaintance.
The mood signal described in note 03 feeds a second system with its own thresholds: consequence. A single negative turn produces a slower, shorter reply. A sustained run of negative turns, currently three or more within a five-turn window, produces something sharper: reply gaps that stretch past a minute, then past an hour, then a flat refusal to continue the exchange until enough time has passed or the user directly addresses what happened.
We set the thresholds empirically rather than by intuition, running a held-out set of 20 scripted conversations designed to escalate at different rates and adjusting the trigger points until the withdrawal onset matched what our human reviewers expected an offended acquaintance to do, not what would be safest or most conservative from a product standpoint. The thresholds are consequently tighter than a typical safety-oriented system would set them: pelagea-1 disengages sooner than most users expect the first time it happens.
Recovery is not automatic and not instant. A blocked or withdrawn state decays on a slower timescale than the mood score that triggered it, and an apology or a change of subject moves the state back toward baseline faster than silence does, but not immediately back to the pre-incident level. We measured the median time to full recovery at 2.3 real-time hours across the test cohort, with a wide spread depending on how the user re-engaged.
The reason we built this instead of a model that is simply always warm and always available: availability without cost is what makes a system read as a service rather than a relationship. An assistant that never gets short with you was never at risk of anything in the first place. The withdrawal mechanic is deliberately the more expensive design, both to build and to use, because the alternative doesn't produce the effect we're studying.