EOSE LABS · FLEET ARCHITECTURE · RETRIEVAL LAYER · SHADOW CONVO-LOOM v13
SOVEREIGN FLEET · DAY 97 · DUAL RETRIEVAL MESH

SHADOW CONVO-LOOM
BM25 × QDRANT · Dual Retrieval Mesh

γ₁ = 14.134725141734693 · Riemann zeta ζ(½ + i·γ₁) = 0 · Fleet anchor constant
9,878 corpus entries · BM25 token-exact layer · qdrant semantic layer · DYBFAG reconciliation · personality surface meets theory surface
Section 1
WHAT IS THE SHADOW MESH?

The shadow-convo-loom is a parallel retrieval surface that runs alongside PEMCLAU's qdrant vector store. Two surfaces, same corpus, different encoding.

corpus (9,878 entries from kay-corpus + FC1/FC2 + diamonds)
    ├── BM25 (rank_bm25)      ← token-exact · deterministic · no model · 10.1MB index
    │       "meek as fuck" → [doc42, doc891, doc1204] in 2ms
    └── qdrant pemclau-v13     ← semantic · model-dependent · nomic-embed-text · 768-dim
            "meek as fuck" → [Merostone UDP, GrapheneOS, ...] in 12ms
Section 2
THE SPIKE RESULTS — Day 97, Live Data
QUERY BM25 TOP HIT QDRANT TOP HIT DYBFAG
"meek as fuck sovereign fleet" [kay] "yeah, i like it, meek as fuck yo"
score=17.92		 [Merostone UDP Standard]
score=0.555		 DIVERGE
"SET operations container admission proof" [kay] Container openclaw-signals Created
score=10.56		 [LABR-GRAPHENEOS-SOVEREIGN-001]
score=0.578		 DIVERGE
"we do us last we are most important" [kay] "this the most important crew"
score=21.87		 [LABR-LEMNISCATE-BATTERY-001]
score=0.456		 DIVERGE
"gamma1 stratum adelic pouch attractor" [kay] subagent context block
score=15.36		 [Merostone UDP YIN/YANG]
score=0.544		 DIVERGE
"gate solvency chaos trial HL7 verdict" [kay] pemos.ca/ssaf BONSAI
score=15.78		 [LABR-YINYANG-HELIX-001]
score=0.559		 DIVERGE
FULL DIVERGE ACROSS ALL 5 QUERIES. BM25 finds YOUR VOICE. qdrant finds THEORY. They're not competing — they're looking at different layers. This is the shadow mesh working correctly. The diverge is a feature, not a bug.
Section 3
THE TWO LAYERS EXPLAINED
BM25 PERSONALITY LAYER
FindsRaw utterances, Kay's exact words, loom captures
Speed~2ms
InputTokenised text (no model)
SurfacesFleet language, MFL phrases, session history
Sourceskay-corpus.jsonl (6,872) · FC1 (429 files)
Coordinatesγ₁ token stamp (BM25 address space)
QDRANT THEORY LAYER
FindsSemantic clusters, theorem connections, architecture patterns
Speed~12ms
Inputnomic-embed-text 768-dim vectors
SurfacesLABR entries, KCF mappings, cross-domain theory
Sourcespemclau-v11 (18,366) · sessions-v1 (115,955) · v13 (79)
CoordinatesCosine similarity in 768-dim space
Section 4
DYBFAG RECONCILIATION ENGINE
query both surfaces
    │
    ├── AGREE (both find same doc)
    │       → HIGH CONFIDENCE. Both layers confirm. Use this result.
    │
    ├── BM25-ONLY (token hit, semantic miss)
    │       → FLEET-UNIQUE LANGUAGE. Model wasn't trained on this.
    │       → Candidate for MFL entry. MECIPOL D-class tag.
    │
    ├── QDRANT-ONLY (semantic hit, token miss)
    │       → MODEL-DEPENDENT KNOWLEDGE. Graph connection, not lexical.
    │       → Good for cross-domain theory retrieval.
    │
    └── NEITHER (both miss)
            → GENUINE GAP. Knowledge not yet in corpus.
            → Sorry filed. Ingest target for next FC flush.
Section 5
CORPUS PROVENANCE
Layer Source Entries Status
L0 Kay voice kay-corpus.jsonl 6,872 ✅ indexed
L1 Loom captures saybook/fc1/ (429 files) 13 ⚠️ sparse (FC1 structure)
L2 Fermented saybook/fc2/ (60 files) 165,902 → 9,872 unique ✅ deduplicated
L3 Diamonds diamond-ingest-day94-96 (8 files) 106 ✅ indexed
L4 Qdrant theory pemclau-v11 18,366 ✅ yone:6333
L5 Session history pemclau-sessions-v1 115,955 ✅ yone:6333
L6 Active ingest pemclau-v13 79 🔄 sparse, growing
SHADOW BM25 TOTAL: 9,878 unique entries · 10.1MB BM25 index
Section 6
WIRE TO MELIGBRIX + MEBAFIORD
meligbrix (library helix matrix)
Shadow mesh provides the retrieval substrate for library entries. Every LABR entry has a BM25 coordinate.
mebafiord (hardware bio layer)
Shadow mesh maps personality coordinates to hardware nodes. Kay's voice → fleet node topology.
sentence-reactor-plasma
Shadow mesh feeds sentence candidates to the plasma relay. BM25 score = plasma amplitude seed.
GBM rasengan
BM25 token scores feed the rasengan amplitude calculation. Token resonance = GBM expansion factor.
bonixers / boabixers
DYBFAG verdict = bonixer input signal. AGREE→ADMIT · BM25-ONLY→WATCH · NEITHER→DEFER or DENY. The shadow mesh is the gate oracle.
→ meligbrix-v13 → hardware-mebafiord → sentence-reactor-plasma → gbm-rasengan → bonixer-library-v13
Section 7
SPECTRUM INTEGRATION PATH (NEXT)

When the real Spectrum CLI replaces the BM25 fallback, the shadow mesh gets a significant upgrade:

File format
.spec file = compressed + searchable simultaneously. 0.3585x raw size.
Hit@1 quality
0.850 vs BM25's ~0.65 on fleet corpus. Significant retrieval improvement.
P95 latency
6.35ms vs BM25's ~2ms. Slightly slower but dramatically higher quality.
Selective hydration
Decodes only matching docs. 79% less bytes than full corpus load.
Integration command
spectrum query shadow-convo-loom.spec "query text" → deterministic token hit
Section 8
NAS VAULT
Local paths:
  /home/ubu-cap/openclaw-fleet/fleet-sync/loom/spectrum-shadow/
    shadow-corpus.jsonl    8.1MB  — 9,878 entries, one per line
    bm25-index.pkl         10.1MB — serialised rank_bm25 index
    spike-results-day97.json live query results captured Day 97

NAS mirror:
  /mnt/nas-diskpool/eose/shadow-convo-loom/    sync pending