Two-device split
Pi 5 is the senses. Desktop/laptop brain is the source of truth for memory, cognition, LLM articulation, policy, learning, and governance.
System reference
The public architecture map.
Condensed from the dashboard System Reference: architecture, flows, truth boundaries, persistence, events, and mode behavior without exposing internal live dashboard state.
Tri-layer cognition
Symbolic truth remains inspectable, neural intuition learns patterns, and LLM articulation only speaks from bounded context.
Epistemic stack
Capability gate, attribution ledger, provenance, identity boundary, scene model, contradictions, calibration, graph, quarantine, audit, soul integrity, compaction, and intention truth.
Persistence
State lives under ~/.jarvis with atomic writes for memories, identities, policy experience, goals, eval data, interventions, and intention records.
Event bus
Kernel, memory, perception, conversation, autonomy, self-improvement, hemisphere, skill, calibration, quarantine, world model, and acquisition events are explicitly cataloged.
Operational modes
Gestation, passive, conversational, reflective, focused, sleep, dreaming, and deep_learning each control cadence, memory writes, and allowed cycles.
Hardware adaptive design
Model choices come from VRAM, CPU, and RAM budgets.
The hardware profile resolves GPU tier, CPU tier, STT model, LLM model, TTS engine, model keep-alive, and whether ancillary ML should live on GPU or CPU. This is why the system avoids casual "just use a bigger model" advice.
Open equipment guideSystem flows
The important paths stay inspectable.
VoicePi mic -> binary PCM stream -> wake word -> VAD -> faster-whisper STT -> route/tool/LLM -> Kokoro TTS -> Pi speaker
PerceptionPi vision events -> attention fusion -> identity resolution -> world model -> conversation and background cognition
Memoryremember() -> quarantine soft gate -> salience advisory -> storage -> keyword/vector index -> MEMORY_WRITE event
Acquisitionclassify -> evidence grounding -> documentation/research -> plan -> review -> implementation -> verification -> registration
Self-improvementscanner -> dedupe -> code generation -> sandbox -> approval -> atomic apply -> health monitor -> rollback
Truthclaim -> capability gate -> attribution -> provenance -> contradictions -> calibration -> graph -> quarantine -> audit