v1.0.0 — Context Spine (ReaderLM + Qwopus + Jina v5)

RISC Thought Engine — Context Spine v1.0

2.4 MB replaces 54 GB. 277 reasoning queries/sec. Zero GPU.

What's Inside

Component	Purpose	Size	Speed
ReaderLM-v2 codebook	Fast lexical routing + HTML detection	425 KB	5,676 q/s
Jina v5 codebook	Embedding anchor (forward pass ground truth)	401 KB	5,676 q/s
Qwopus 27B gates	Deep semantic context (8 of 64 layers)	2 MB	277 ctx/s

Architecture

spider-rs → raw HTML → ReaderLM-v2 (candle, BF16) → clean markdown
  → Qwen tokenizer (151936 vocab) → codebook → centroids
  → F32ThinkingEngine (softmax T=0.01) → peaks
  → Qwopus gate EKG (8 layers × 4 roles) → deep context
  → AriGraph triplet_graph → SPO knowledge
  → NARS truth revision → confidence
  → ContrastiveLearner → table improves over time

Proven Results

• HighHeelBGZ i16 encoding: 100% top-5 fidelity (7,127× compression)
• Softmax T=0.01: 100% agreement with plain cosine (attractor collapse SOLVED)
• Gate EKG: perfect discrimination (0/8 agreement between different topics)
• False triplets: correctly detected as LOW confidence
• Garbage detection: entropy < 1.0 = bad ReaderLM output

Railway Deployment

# Clone + deploy
git clone https://github.com/AdaWorldAPI/lance-graph.git
cd lance-graph
railway up

Or use the Dockerfile directly:

docker build -f crates/thinking-engine/Dockerfile.railway -t thinking-engine .
docker run -p 8080:8080 thinking-engine

Download

curl -LO https://github.com/AdaWorldAPI/lance-graph/releases/download/v1.0.0-context-spine/context-spine-v1.0.tar.gz
tar xzf context-spine-v1.0.tar.gz
# 4.2 MB extracted: codebooks + manifest.json

v0.3.0 — HighHeelBGZ 256 + 4096 Centroids

HighHeelBGZ Codebooks: 256 + 4096 Centroids (Jina v5)

Both centroid sizes for the thinking engine, from Jina v5 0.6B (1024D) safetensors.

Size Comparison

K	i8 table	codebook idx	Total	vs GGUF (1.2 GB)
256	64 KB	297 KB	368 KB	3,248× smaller
4096	16 MB	297 KB	16.3 MB	70× smaller

Quality (softmax T=0.1, 10 cycles)

K	Cosine Range	Mean	Note
256	[-0.19, 0.68]	0.224	Dense MatVec works (proven 70-78% on Qwen3-VL)
4096	[-0.50, 0.87]	0.166	Wider range, needs sparse top-K graph for MatVec

Files per tarball

cosine_matrix_KxK.f32       Raw ground truth
distance_table_KxK.i8       HighHeelBGZ i8 (LOSSLESS, proven)
distance_table_KxK.bf16     HighHeelBGZ BF16 (LOSSLESS, proven)
codebook_index.u16          Token → centroid assignments
encoding_metadata.json      Calibration params

Usage

# Download
curl -LO https://github.com/AdaWorldAPI/lance-graph/releases/download/v0.3.0-highheelbgz-256-4096/jina-v5-256.tar.gz
curl -LO https://github.com/AdaWorldAPI/lance-graph/releases/download/v0.3.0-highheelbgz-256-4096/jina-v5-4096.tar.gz

Sparse Branch Graphs (included in 4096 tarball)

K	Edges	Size	T=0.01 Top-5	Diversity
8	32K	256 KB	32%	18/20
16	65K	512 KB	30%	19/20
32	131K	1024 KB	31%	18/20

Dense 256 at T=0.01 = 100% top-5, 20/20 diversity, H=0.000 (PERFECT).

Files: branch_graph_4096xK.indices.i32 + branch_graph_4096xK.values.f32

v0.2.0 — 7-Lane HighHeelBGZ Codebooks

7-Lane HighHeelBGZ Codebooks

Two embedding models, 7 encoding lanes each (770 KB per model).

Models

Model	Params	Hidden	Cosine Range	Gap
Qwen3-VL-Embedding-2B	2B	2048D	[-0.846, 0.539]	0.132
Jina v5	0.6B	1024D	[-0.187, 0.675]	0.128

7 Lanes

1. u8 CDF (percentile rank)
2. i8 direct (round(cos*127), signs preserved)
3. u8 γ+φ (golden ratio redistribution)
4. i8 γ+φ signed
5. f32 SiLU correction deltas
6. BF16 direct (StackedN source precision)
7. u8 spiral drift (highheelbgz reconstruction, stride=11)

Dockerfile Usage

# Download and extract
ADD https://github.com/AdaWorldAPI/lance-graph/releases/download/v0.2.0-7lane-codebooks/qwen3-vl-embedding-7lane.tar.gz /app/codebooks/
ADD https://github.com/AdaWorldAPI/lance-graph/releases/download/v0.2.0-7lane-codebooks/jina-v5-7lane.tar.gz /app/codebooks/

Encoder Command

cargo run --release --features calibration --example seven_lane_encoder -- qwen3-vl-embedding

Branch: claude/risc-thought-engine-TCZw7

Qwopus 27B per-layer tables + codebooks + sixpack

Per-layer distance tables for Qwopus 27B (64 layers × 5 roles).

Includes:

• 64 layer directories with attn_qkv, ffn_gate, ffn_up, ffn_up_silu, ffn_down tables (256×256 u8)
• Token embedding tables (256×256) and assignments (248K tokens)
• Gate comparison pair (raw vs SiLU-corrected)
• Tokenizer (Qwen2 BPE, 7 MB)
• layer_stats.json
• Sixpack layer 20 (Gate/Up/Down separated)
• 64×64 codebook tables (8 models)

Total: ~35 MB. Streamed from 53.8 GB BF16 GGUF.
Place in: crates/thinking-engine/data/

Tokenizer files for 6 models

HuggingFace tokenizer.json files for cross-model calibration.

Files:

• jina-v3-tokenizer.json (8.7 MB, XLM-RoBERTa 250K vocab)
• bge-m3-tokenizer.json (8.7 MB, XLM-RoBERTa 250K vocab, same as Jina v3)
• jina-v5-tokenizer.json (11.4 MB, Qwen3 151K vocab)
• xlm-roberta-de-tokenizer.json (8.7 MB, German NER variant)

Place in: crates/thinking-engine/data/
Or: tokenizer_registry.rs downloads via from_pretrained() automatically.

Sources:

• XLM-RoBERTa: FacebookAI/xlm-roberta-large-finetuned-conll03-english
• Qwen3: jinaai/jina-embeddings-v5-text-small-text-matching
• Qwen2 (Reranker/Qwopus): already in Qwopus data dir (7 MB)

Qwopus 27B 4096-centroid model data

4096 CLAM centroids + distance table for Qwopus 27B token embeddings.

Streamed from 53.8 GB BF16 GGUF via HTTP range requests in 148s.

Files:

• token_embd_centroids_4096x5120.f32 (80 MB) — 4096 centroid vectors
• token_embd_4096x4096.u8 (16 MB) — HDR CDF distance table

Place in: crates/thinking-engine/data/Qwopus3.5-27B-v3-BF16-silu/

Built with: reservoir sampling 16K from 248K tokens → CLAM furthest-point → 4096 centroids
Cluster balance: min=1, max=535, mean=61 tokens/cluster

bgz7 model indexes

Pre-built bgz7 indexes for 5 Qwen3.5 models (685 MB total).

Use: cargo run --manifest-path crates/bgz-tensor/Cargo.toml --features hydrate --bin hydrate -- --download MODEL