Mood as a control signal

Every human turn is scored on a continuous scale from -1 to 1 before the model replies. The score is not exposed to the user and is not a sentiment label attached after the fact for analytics; it is read by the response pipeline before generation and used to set three things directly: how long the model waits before answering, how many words it uses, and how likely it is to volunteer information nobody asked for.

We validated the scoring model against the same 12.4k-turn corpus used in note 04, this time with three annotators independently rating the emotional trajectory of each exchange and resolving disagreement by majority. Pelagea-1's own scores correlated with the human-annotated trajectory at r=0.81, with the largest divergence in exchanges that shifted tone abruptly inside a single turn, which the current model tends to average rather than catch immediately.

The behavioural routing is where this stops being sentiment analysis and starts being a control system. A mood reading at or below -0.3 measurably lengthens reply latency and shortens replies in the same session, matching the cooling pattern shown in our live demo, where a dismissive remark drops a warm exchange from replies in the 200ms range to a terse response after 1.4s. A reading at or above 0.4 does the opposite: latency drops and the model is more likely to add unprompted detail, the two turns after a positive spike being 34% longer on average than the baseline for that user.

The finding we did not expect going in: staying level through a hostile exchange, rather than mirroring it, was rated by blind reviewers as the single strongest signal that they were talking to a machine. People escalate. A system that tracks the escalation accurately but declines to participate in it reads as controlled in a way no human conversational partner reliably is, which means the most human-sounding failure mode is actually the more volatile one.