train-convo-steering (Skill v2)

Voice-first runtime steering + nightly deep analysis to learn per-user conversation steering priors.

This skill is designed for:

• Live voice: per-turn inference must be fast and bounded (preset selection, not multi-candidate judging).
• Nightly deep learning: optional DeepSeek V3 (TEE) judge calls to improve labels and update per-user priors.

Concepts

Collaboration State

A compact state bucket per turn (tempo, trust, alignment, affect, control) mapped to {low, mid, high}.

Steering Presets

Configuration of response knobs (length, questions, initiative, certainty, grounding) and voice prosody.