LIVING GRAPH STRUCTURE
NODES = TASKS + PATTERNS
Every task attempted across all crew sessions is a node. Type: code / proof / build / hunt / audit / wave-search. Nodes colored by wave activation.
EDGES = CAUSALITY
Which conversation → which code → which fix → which sorry closed. Causal edges γ₁-timestamped. 2-hop GraphRAG finds the chain even across sessions.
Y1 STRATUM HELIX
yone owns the living graph substrate. γ₁ anchor. PEMCLAU pemclau-v11 (18,357 pts). nomic-embed-text on yone:11434. All local fleet interactions flow through yone's stratum.
PLASMA RELAY LOOP
TUI session → wave reactor classify → embed → PEMCLAU upsert → next query is smarter. Runs in ~2s per interaction. Session becomes non-amnesiac within the same session.
CORPUS STATS
kay-corpus.jsonl: 6,872 entries
diamond-day94: 8 diamonds
diamond-day95: 11 diamonds (+4 math)
PEMCLAU total: 18,357 pts
Wave classifications: 18 waves
Graph edges: 80,979+
HARBOR TUI LINK
Terminal Bench 2.0: 89 tasks
Deep Agents: 42.5% (= Claude Code)
Our fleet: k3d+Docker sandbox
Each silo = isolated Harbor env
→ TUI runs TB task → result → PEMCLAU
→ Next run: smarter from prior result
LEARN / RELEARN
Every TUI interaction = new node
Success → W14 SEAL edge
Failure → W7 INVERSION edge (try flip)
120Hz RL clock = learning rate
γ₁ zero positions = the non-amnesiac checkpoints