Architecture Hub

A system blueprint, not a list of feature cards.

The architecture page now shows nodes, interfaces, evaluations, and risk boundaries so each layer has a clear purpose and a clear edge.

Moonweave system graph

Architecture Atlas

Layer interfaces

Maintains recognizable identity, conversational style, preferences, and allowed modes of agency.

Stores, retrieves, edits, and deletes long-horizon context with source records and user control.

Inputs: Conversation events / Artifacts / Correction requests
Outputs: Recall packets / Graph links / Retention records
Evaluation: Recall precision, deletion compliance, source traceability, and hallucination resistance.

Tracks bounded affective state so emotional continuity remains inspectable rather than decorative.

Inputs: Dialogue tone / User feedback / Persona constraints
Outputs: Affect state / Tone modulation / Stress flags
Evaluation: Trust calibration, non-manipulation, and state stability across sessions.

Connects language, visual context, synthetic voice, and virtual expression in one interaction loop.

Inputs: Text / Audio / Visual context / Environment events
Outputs: Responses / Voice cues / Avatar directives / Tool calls
Evaluation: Latency, disclosure clarity, modality grounding, and turn-taking quality.

Translates internal state into supervised virtual and physical action pathways.

Inputs: Runtime plans / Sensor state / Human approval
Outputs: Avatar motion / Simulation actions / Robot API requests
Evaluation: Sandbox compliance, reversibility, human override, and action safety.

Keeps media, data, model behavior, and embodiment inside public rights-aware boundaries.

Inputs: Logs / Media metadata / Risk reports
Outputs: Release gates / Audit views / Policy constraints
Evaluation: Incident readiness, traceability, accessibility, and policy coverage.

Risk boundaries

Persona

Identity overfitting, false intimacy, and unsafe autonomy claims.

Memory

Over-retention, private-data leakage, and unsupported relationship claims.

Affect

Emotional manipulation, dependency loops, and hidden persuasion.

Runtime

Unlabeled synthetic media, likeness confusion, and modality hallucination.

Embodiment

Unsafe physical action, authority confusion, and poor emergency-stop design.

Safety

Ambiguous ownership, undisclosed generation, and unreviewed capability release.