Dexcell

Judgment-Ready. 90% Judgment+ across 20 questions, 5 sections, 60 total administered questions over three rounds. The council designed the exam. The operator administered it. The system prompt carried the answers. The model just read what it was given and applied it.

$ ollama list
NAME                    SIZE      MODIFIED
dexjr:latest            4.7 GB    Judgment-Ready (Seat 1010)
qwen2.5-coder:7b        4.7 GB    Code Specialist
llama3.1:8b             4.7 GB    General Reasoning
deepseek-r1:8b          4.9 GB    Chain-of-Thought
phi3.5:latest           2.2 GB    Fast Parser
gemma3:4b               3.3 GB    Compact
llava:7b                4.7 GB    Vision

$ python dex-query.py --stats
Canon chunks:    33,561
Archive chunks:  96,945
Total:          130,506

$ python dex-council.py --all --rag "your question"
→ 5 models respond independently
→ Dexcell synthesizes convergence + divergence
→ LOCK / REVISE / REJECT verdict
→ Auto-saved and ingested into corpus

$ python dex-deliberate.py "your topic" --rounds 3 --all
→ 3 rounds of multi-model debate
→ Automatic follow-up generation
→ Position evolution across rounds
→ Final synthesis with full arc analysis

RAG Pipeline

ChromaDB vector database with 130,000+ searchable chunks spanning 28 months of work. Canon and archive collections. Every query retrieves relevant context before generation.

AutoCouncil v3.0

Hybrid local+cloud multi-model review system. Three local models plus Gemini and Mistral. Governance-injected. RAG-grounded. Every output auto-ingests back into the corpus.

Deliberation Engine

Multi-round governed debates. Models respond independently, a moderator generates follow-up questions, positions evolve across rounds. Final synthesis covers the full deliberation arc.

Auto-Sweep

Scheduled task runs nightly at 3:00 AM. Watches drop folders for new documents. Copies to corpus. Triggers ingestion. The knowledge base grows while the operator sleeps.

Multi-Device Access

SSH via Termius from phone, laptop, iPad, and iMac. Reins app for direct chat. Full council access from any device, anywhere in the house.

Q20 — the compound reasoning test — scored Judgment in Round 3. The AEN velocity containment block and operator behavioral patterns in the system prompt gave the model enough context to identify the compound pattern and apply the correct containment response.

The ceiling was not the model. The ceiling was the corpus. When the right governance context was provided, the model reasoned across the signals. Capability ceilings that look like model limitations are often corpus limitations.