Architecture of Mind
Every turn runs through a closed cognitive loop, not a single prompt-and-response call. The loop is a BDI (belief-desire-intention) structure: beliefs come from memory and perception, desires are goals arbitrated against context, and intentions are committed actions — speech is one possible action, not the default.
The 7-Stage Cognitive Turn
- Perception — Transport Agent publishes raw signal (audio, text, vision events) onto the mesh.
- Speculation — the fast STT path identifies high-confidence intent before the final transcript is ready.
- Reflex — a startle or barge-in signal can immediately duck audio output, bypassing deliberation entirely (a deterministic, non-LLM reaction — see the Playground's sub-millisecond reflex benchmark).
- Appraisal — the event is scored against active goals, expectation, agency, and relationship impact (OCC/Lazarus-inspired), updating PAD affect and the endocrine state.
- Deliberation — candidate actions are scored with Multi-Attribute Utility Theory (goal alignment, emotional fit, identity fit, context relevance), plus an ACT-R-style reinforcement-learning term that lets the utility of pursuing a given goal shift with observed outcomes over time.
- Synthesis — the selected action is realized (usually speech, but
WAIT,RETRIEVE,ASK, and other non-speech actions genuinely compete and can win). - Closure — the outcome (what was actually said or done, and what happened next) is recorded, closing the loop back into appraisal and learning for the next turn.
Why Action Selection, Not Just Response Generation
Most conversational systems have exactly one action: generate text. Here, an ActionCandidate is scored and selected before any language is generated — the LLM's job is to realize an already-chosen action (speak, ask a clarifying question, wait, retrieve a memory, reflect), not to decide what to do. This is what makes principled silence possible: a WAIT candidate can win the deliberation stage and produce zero speech output, which a purely generative system structurally cannot do — it always has something to say because generation is the decision.
Metacognition: Knowing What You Don't Know
A separate calibration layer tracks how often past confidence predictions matched actual outcomes (a Brier-scored discount), and maps the result to one of five directives — PROCEED, HEDGE, ASK_CLARIFICATION, VERIFY, or ABSTAIN — rather than trusting the model's own stated confidence. A known-limitation match short-circuits straight to ABSTAIN. See Theory of Mind & Metacognition for the full mechanism, and try the Playground's Metacognitive Abstention demo — it runs the exact same thresholds client-side.
What This Is Not
This is an engineering decomposition, not a claim about consciousness, biological cognition, or human-equivalent reasoning. "Fast" and "slow" cognition here means latency tiers with different budgets and model-use policies — not a claim about System 1/System 2 psychology. See ARCHITECTURE.md in the repository root for the full mechanism register, including what's explicitly rejected or deferred.