System Map

The single diagram of what exists and who talks to whom in SMARTHAUS.

Nine systems. One shared field. Every system is either the memory substrate (RFS), something that routes or orchestrates using that substrate (MAIA, CAIO, TAI, VEE), something that feeds or uses it (NME, VFE), or governance (MGE).

---

System Decomposition

System	Repositories	Role
RFS	rfs-core, ResonantFieldStorage, rfs-sdk	Memory substrate: 4-D resonant field; exact recall + associative resonance + proactive discovery.
MAIA	MAIA	Intent engine: semantic intent resolution, routing to VFE/RFS/marketplace via master equation.
CAIO	caio-core, caio, caio-sdk	Universal AI orchestration; deterministic policy; contract-gated execution.
TAI	tai-core, tai, tai-sdk	Product: voice-first personal assistant; integrates CAIO, VFE, RFS, NME, MAIA, VEE.
AIVA	Archetype specification	Enterprise assistant archetype; triadic AIOS/AQL/AEF architecture; RFS as memory substrate.
VFE	vfe-core, VerbumFieldEngine, vfe-sdk	Inference: model API gateway; OpenAI-compatible; model selection calculus $S_{total}$.
NME	NotaMemoriaEngine	Memory structuring: trait extraction; encoding into RFS via $E(\text{Extract}(\Psi_t))$.
VEE	VoluntasEngine	Intent classification: native C++/pybind11; RL-based intent parsing.
MGE	mge-core, MathematicalGuaranteeEngine, mge-sdk	Mathematical Governance Engine: rules-as-code; compliance API; MA enforcement sidecar.

---

Cross-System Coupling

How state flows between systems:

Producer	Consumes	Emits	Signal Form
RFS	Bytes, embeddings (ingest); query vectors (associate)	Field state $\Psi$; matched-filter response $r = E^H \Psi$; recall bytes; quality metrics $Q$, $\eta$, capacity	Field state; vectors; metrics
MAIA	Query $q$, context $C$, market $M$; RFS resonance $W$	Route decision (VFE / RFS / marketplace)	Discrete choice + scores
NME	Raw user input	Trait structures; encoded waves for RFS	Structured records; $E(\text{Extract}(\Psi_t))$
VFE	Request, model selection weights	$S_{total}$; token stream	Scalar score; tokens
CAIO	Request, policy	Adapter choice; response	Contract-satisfying I/O
VEE	Context (e.g. from RFS)	Intent classification (RL policy)	Label or distribution
MGE	Rules (.mdc, .yaml), request/action	/check result	Pass/fail + reasons
TAI	User input; responses from CAIO, VFE, RFS, NME, MAIA, VEE	Requests to subsystems; final user response	Requests/responses; voice; embeddings

RFS as the Memory Substrate

RFS is the connective tissue. MAIA needs RFS to compute resonance $W(q, C)$ for routing decisions. NME writes trait waves into RFS. TAI uses RFS for both associative and exact recall. VFE may use RFS for context retrieval. CAIO may read RFS metrics for policy decisions.

If RFS were removed, MAIA's resonance term $W$ would be undefined, NME would have nowhere to store trait waves, and TAI would lose its memory. The field is not optional infrastructure — it is the integration substrate.

MAIA as the Router

TAI and other consumers send query + context to MAIA. MAIA's master equation combines intent $I$, alignment $A$, reinforcement learning $RL$, and RFS resonance $W$ to produce a route decision. Replacing this with "call an LLM to decide" removes the same-input-same-route guarantee.

CAIO as the Orchestrator

Requests pass through CAIO for policy enforcement. Contract pass/fail is deterministic — same request and policy always produce the same enforcement decision. CAIO ensures that governance is executable, not advisory.

---

Determinism Boundaries

Not everything in the system is deterministic. The architecture explicitly separates where determinism is enforced from where probabilistic behavior is permitted.

Enforced Determinism (same inputs → same outputs)

• RFS encode/recall (seeded $E$, $\Pi_{assoc}$)
• RFS byte channel (AEAD encryption)
• MGE rule evaluation
• CAIO policy application
• VFE model selection calculus (same request + weights → same $S_{total}$ and model choice)

Permitted Stochasticity

• MAIA RL component (learning evolves policy)
• VEE RL policy learning
• LLM inference via VFE (token sampling)
• NME trait extraction (if model-based)
• TAI inference responses (downstream of LLM sampling)

The design principle: learned or stochastic components operate on top of or beside deterministic ones. MAIA's route uses deterministic $I$, $A$, $W$ terms and a stochastic $RL$ term. VFE's model selection is deterministic; the inference through the selected model is stochastic. Invariants and gates apply to the deterministic surfaces. Probabilistic components are bounded, not eliminated.

---

Why Not LangChain + a Vector Database?

This is a fair question. The answer is structural, not preferential:

• Single field, multiple channels. RFS provides exact recall, vector similarity, interference-aware (corpus-aware) search, and proactive discovery from one state $\Psi$. A vector database stores isolated points. It has no superposition, no $Q$/$\eta$ guardrails, no band separation between exact and associative channels.

• Equations govern behavior. Invariants and lemmas pin capacity, recall error, stability, and resonance to formal bounds. CI pipelines enforce them on every change. LangChain + vector DB do not ship formal operator calculus with CI-bound verification artifacts.

• Intent and routing are math. MAIA's $\text{Route}(\cdot)$ is a defined function of $I$, $A$, $RL$, $W$, where $W$ couples directly to the RFS field. Replacing this with "call an LLM to decide the route" removes same-input-same-route guarantees.

• Memory is structured and trait-aware. NME encodes traits as field contributions $E(\text{Extract}(\Psi_t))$ into the same RFS field. A vector database stores vectors and optionally blobs — no 4-D resonant field with band separation and interference-aware scoring.

• Proactive discovery before any query. RFS can reveal clusters, anomalies, and field topology without a user query. Vector databases are query-reactive only.

• Determinism with rollback. MA defines rollback criteria per invariant. Scorecard gates block release when guarantees fail. There is no equivalent in a generic orchestration + retrieval stack.