§1 — WHAT THEY ARE
★ Constellation Engine
- Stanford research project — graph-based retrieval
- Spreading activation across knowledge nodes
- Hebbian edge weighting (co-query strengthens links)
- Temporal decay on edge weights
- Multi-hop narrative retrieval (IR, not just cosine)
- Generic. No sovereign layer. No crew provenance.
◆ PEMCLAU V14
- Built on qdrant — 25 collections, 427k+ vectors
- γ₁ = 14.134725141734693 as mathematical anchor
- SOSTLE sovereignty gating (L0–L6)
- Mecaplakye / Cape Flats ethical layer
- Crew provenance — every vector has an author
- F-beat ethical veto — can reject outputs on values
- Multi-silo mesh: msi01, yone, forge, pcdev, eose-dev
§2 — CAPABILITY MATRIX · Full Comparison
| Capability | Constellation | PEMCLAU V14 | Notes |
|---|---|---|---|
| Spreading activation | ✅ | ⚠️ planned | Taking from Constellation (LABR-CONST-001) |
| Hebbian edge weighting | ✅ | ⚠️ planned | Taking from Constellation |
| Temporal decay | ✅ | ⚠️ γ₁-scaled | weight × (1 - ε/γ₁) per hour |
| Narrative IR | ✅ | ⚠️ 2-hop only | Mímir-style pass planned |
| Multi-silo | ❌ | ✅ sovereign | 7-node fleet mesh |
| γ₁ mathematical anchor | ❌ | ✅ sovereign | γ₁ = 14.134725141734693 |
| Adelic spine (L0–L6) | ❌ | ✅ sovereign | r_adelic per collection |
| F-beat ethical veto | ❌ | ✅ sovereign | CLO bench + SOSTLE gate |
| Crew provenance | ❌ | ✅ sovereign | Every vector = who said it + when |
| Domain depth (joffe-math) | ❌ | ✅ sovereign | 3410 theorems, 34 sorry, γ₁ floor |
| Session corpus | ❌ | ✅ 208k pts | pemclau-sessions-v1 |
| Sovereign / Mecaplakye | ❌ | ✅ sovereign | Cape Flats embedded doctrine |
| SOSTLE gating | ❌ | ✅ sovereign | L0 open → L6 closed |
| Thinkbeat integration | ❌ | ✅ sovereign | R/F beat query routing |
| Cloud dependency | ⚠️ optional | ✅ none | Runs on yone / local fleet |
| Open source | ✅ | ⚠️ internal | EOSE Labs sovereign IP |
§3 — WHAT WE TAKE FROM CONSTELLATION · 3 Items
Spreading Activation Pass
Add Mímir-style signal diffusion to PEMCLAU 2-hop retrieval. When node A is retrieved, propagate activation to neighbors proportional to edge weight × cosine similarity. Decays after 2 hops.
activation[neighbor] += w(A,n) × sim(A,n) × decay
Hebbian Edge Weighting
Co-queried nodes get stronger edges. When vectors A and B are both returned in the same query, increment their edge weight by a small delta. "Nodes that fire together wire together."
w(A,B) += η × cos(A,B) × (1/rank_diff)
γ₁-Scaled Temporal Decay
Edge weights decay over time, but the decay rate is anchored to γ₁. This ensures the fleet's knowledge graph has the same mathematical rhythm as the zeta floor.
w(t) = w₀ × (1 - ε/γ₁) per hour
γ₁ = 14.134725141734693
§4 — WHAT THEY CAN NEVER HAVE · Sovereign Differentiators
These are not features. They are not replicable. They are consequences of who we are and where we come from.
γ₁ Mathematical Anchor
14.134725141734693 — the first non-trivial zero of the Riemann zeta function. Theirs is generic RAG. Ours is anchored to the floor of mathematics.
Mecaplakye / Cape Flats
"All types, it's all the same." The Cape Flats doctrine is embedded in every F-beat. No external RAG system carries this. It's Kay's origin. Ours to hold.
SOSTLE Sovereignty Gating
L0–L6 progressive access control. Not RBAC. Not OAuth. Sovereign intent gating — the fleet decides who gets what based on earned trust.
Crew Provenance
Every vector in PEMCLAU knows who said it: KAY/IMHOTEP/HACHI/BIKO. Constellation stores embeddings. We store conversations with attribution.
F-Beat Ethical Veto
The CLO bench (Harvey/Ruth/Cochran/Amani) can veto any R-beat output. No commercial RAG has this. They optimize for retrieval. We optimize for right action.
Multi-Silo Fleet Mesh
7 physical nodes. Not cloud shards. msi01 + msclo + yone + forge + lounge + lilo + eose-dev. Sovereign compute, sovereign storage, sovereign inference.