Snapkitty/sovereign-chat-harness

SnapKitty Sovereign Chat Harness

If it can't prove it, it won't say it.

Every response is checked by five deterministic ERE gates and sealed with a SHA-256 audit receipt before you see it. Responses that fail are blocked and explained — not silently returned.

How it works

prompt
  │
  ▼
P1 verify(prompt)     — MetaSum coherence check on prompt
  │
  ▼
LLM (Ollama local / HF Inference)
  │
  ▼
P2 verify(response)   — MetaSum coherence check on response
  │
  ▼
P3 curate(response)   — Dream Cycle recovery gate
  │
  ▼
P4 prove(joint)       — entropy invariant ≤ 0.20
  │
  ▼
P5 seal()             — WORM SHA-256 receipt
  │
  ▼
allowed? → show response + proof panel
blocked? → show why, suppress response

Gate table

Gate	External label	What it checks
P1	verify	Prompt phase coherence: `
P2	verify	Response MetaSum: model output has sufficient signal
P3	curate	Dream Cycle: coherence restored after low-MetaSum event
P4	prove	Joint entropy ≤ 0.20 — Weil-bound invariant
P5	seal	WORM SHA-256 receipt for audit chain

Why it doesn't hallucinate

The coherence threshold τ = N/2 = 512 enforces that the MetaSum magnitude of any response exceeds the Weyl bound √(N·log Q) ≈ 89. Responses below this bound are recovered via Dream Cycle or blocked entirely.

Entropy invariant H ≤ 0.20 is derived from the same bound: post-Dream-Cycle entropy drops to ≈ 0.07, well inside the safe zone.

Local setup

git clone https://github.com/SNAPKITTYWEST/sovereign-chat-harness
cd sovereign-chat-harness
pip install -r requirements.txt
cp .env.example .env
# edit .env — set HF_TOKEN and/or start Ollama

# With Ollama (recommended):
ollama run snapkitty-merged  # or any GGUF model

# Without Ollama (HF Inference):
export HF_TOKEN=hf_...

python app.py
# open http://localhost:7860

RTX 3080 Docker (8 GB VRAM)

See Dockerfile — one-command local sovereign stack with audit trail:

docker build -t snapkitty-chat .
docker run -p 7860:7860 --gpus device=0 \
  -e HF_TOKEN=$HF_TOKEN \
  snapkitty-chat

Every response prints a WORM hash. NVIDIA can't offer this.

Terminology

Internal terms (code) map to external terms (README/tags/HF search):

Internal	HF search term
FSM pipeline	data curation / filtering
ERE gates P1-P5	verification / guardrails
WORM sealing	reproducible audit log
SovMonster / Lean proof	formal verification
SoulVM JIT	local inference runtime
Dream Cycle	hallucination recovery
MetaSum	coherence scoring

Models

• Ollama (local, preferred): snapkitty-merged Q4_K_M GGUF, RTX 3080, 8 GB VRAM
• HF Inference (cloud): Snapkitty/snapkitty-merged
• Any Ollama-compatible model works — swap via OLLAMA_MODEL env var

License

BSL-1.1 / AGPL-3.0 / MPL-2.0 — Patent Pending, BEL ESPRIT D ACCORD TRUST HOLDINGS INC.

💼 Commercial License

Snapkitty code is free and open under AGPL-3.0 for open-source use. Building a commercial product or service? A proprietary commercial license from Snapkitty Collective LLC lets you ship this code without the AGPL's source-sharing and network-use obligations.

→ Get a commercial license · A.parr@belespritdaccord.uk