Two-layer schema: 16 immutable core fields (Layer 1 — the Meridian protocol) + 20+ extensible fields (Layer 2 — personal only, null defaults, never emitted). Layer 1 never changes. Layer 2 grows without migration.

BGE-M3 (1024-dim) as Meridian standard. Rob must migrate 16,717 holons from 768 → 1024-dim. Recommend batch re-embed (~4h on CPU).

If gravity = retrieval frequency → rich-get-richer. If gravity = validation count → "validation" undefined. If gravity = LLM-assigned → only as reliable as extraction. The entire compounding proposition depends on this.

SQLite table: confidence_history (principle_id, old_score, new_score, timestamp, trigger, context). Principles dropping >0.3 from peak flagged for review. Storage: ~34K rows/year. Negligible.

NODE_SCHEMA with node_type='error'. Fields: error_type, related_principle_id, context, outcome, severity. Embeds on error description for similarity search.

Per-domain threshold. Current 0.85 validated on marketing/psychology. Health may need 0.82, finance may need 0.88. Test across all three founders' domain data.

Whitepaper v5 Section 05.5, Risk Register CON 03.

"Interest rate increases slow capital allocation to emerging markets" belongs in macro_economics, equity_analysis, market_structure, AND geopolitics simultaneously. Single-label fails.

Ingestion writes: (1) vector to LanceDB, (2) metadata to SQLite, (3) edges to graph. If step 1 succeeds and step 2 fails = ghost record. LanceDB is eventually consistent, SQLite is ACID, graph is in-memory.

Clean agent loop: message → context assembly (SPINE + RAM + retrieved knowledge) → LLM call → tool execution → response → state update. Tools registered via config. Context assembly pluggable per domain.

SPINE > Beliefs > RAM. SPINE overrides beliefs. Beliefs override RAM.

Unified interface: call(messages, model_key, tools=None). Provider by config. Cascade logic optional. Will v4 adds: router.py as single entry point with decompose() for subtask routing. Task types: EXTRACTION(8B), REASONING(30B+), SYNTHESIS(30B+), FO

SQLite table for MVP. Named channels, priority levels, structured messages. Polling-based. Upgrade to Redis later if latency matters.

Structured YAML. Three enforcement layers: (1) system prompt (probabilistic), (2) output validator (deterministic), (3) tool permissions (structural).

Whitepaper v5 OQ-01. Q is building this now with Metatekt.

Core pipeline: source → chunk → extract → embed → store. Config-driven. No hardcoded references.

Image → vision model → text → standard pipeline. All modalities produce NODE_SCHEMA output.

Filesystem watcher (watchdog) on /inbox/. Webhook endpoints optional. New file → detect type → route → ingest → harden → notify.

Before batch: snapshot → run against staging → compare (connectivity, gravity, duplicates) → rollback if regression. LanceDB versioning (native time-travel) may eliminate storage cost.

Schema check + dimension check + signature verify + anomaly scan + rollback guard.

Stage 2 asks LLM to produce atomic principles with 16 fields. No verification between extraction and writing. Atomicity is not well-defined for LLMs. Three models produce structurally different outputs from the same paragraph.

Large chunks (2,048 tokens) for Stage 2 extraction. Small chunks (256 tokens) indexed for retrieval scoring. Extract principles from large chunks, link them to small chunks for precision. Extraction chunk size and retrieval chunk size are separate co

AES-256. OS-level default + VeraCrypt guide for paranoid clients. Hardware key to unlock.

All inference, embeddings, storage local. Internet opt-in only. Validated: GHOSTNET runs air-gapped on Raspberry Pi.

PII detection + content classification + domain relevance + output redaction. All external inputs sanitized before touching KB.

heartbeat.json updated every 60s. Recovery daemon triggers restart on 5min silence.

Physical. Pull ethernet. Power down. No remote override. Design principle, not feature.

retrieval_log table: query_hash, retrieved_record_ids, records_used_in_response (boolean per record). Monthly retrieval quality reports from day 1. This is how you answer: "is this system working?"

HyDE (Hypothetical Document Embeddings): generate a hypothetical answer to the query locally → embed THAT for retrieval instead of the question. The hypothetical lives in the same semantic space as stored principles. Privacy preserved: generated loca

Open WebUI + custom LanceDB RAG plugin for MVP. Dashboard primary, chat secondary (Will v4 Principle 07).

WhisperX local. Validated in production.

Defer to post-MVP. CLI + chat for founding operators. Dashboard is retention, not acquisition.

Whitepaper v5 OQ-04, founders call.

Idle-period processing. Random sample N principles, cross-domain connections, store in dreams.lance. Nice-to-have for MVP, critical for v2.

Supervisor + specialized workers. Post-MVP. Base model ships with single agent.

System proposes, user confirms. SQLite table. Gradient of control (Q + Will): user-selectable autonomy from full approval to fully hands-off.

Emission packet with principle, confidence, gravity, domain, validation_count, signature. Simple signature for founders, ZKP at 33+ nodes. Codex auto-creation (Q + Will): personal AI detects when research hits contributional level, auto-creates codex

Where does Mother run? Shared VPS recommended for MVP.

Statistical anomaly detection + minimum validation threshold + cross-validation (≥2 nodes) + audit trail. Gradual drift detection over time windows for the hardest attack vector.

Whitepaper v5 OQ-03.

Shared table with agent_id column + filtered views. Each agent sees only its own rows. Queryable by time, success/failure, manifesto alignment.

MANIFESTO.md per agent. Mission, capabilities, aspirations, growth metrics, gaps. Evolves with user + agent dreams + external disruptions + personality shifts + family expansion.

dream_cycle(agent_id) → activity log + manifesto + KB query → LLM generates mutations[], dream_log, manifesto_update. Rob's dream mechanism + Q's manifesto concept = dreaming from past to future.

Three types: user_mutation (auto-propagate), dream_mutation (approval queue), collective_mutation (from Mother). Explicit command for MVP, inferred from patterns later.

Interviews client → proposes agent config → creates SPINE + MANIFESTO + beliefs per agent → initializes logs → self-destructs. The most important onboarding experience. See also SPEC-046 (Seed as DNA).

Extract principles from dream output, store as node_type='dream_insight', source_id=agent_id. Start at confidence 0.5. Mini-hardening after.

Receive anonymized dream_insights, cluster by similarity, synthesize innovations, broadcast as collective_mutations. Manual at 3 founders, automated at 33+.