Virtue-based AI coherence network for local LLMs.
5 agents. 6 virtues. One attractor: Q∞ = 0.83929
q3sh is a framework for running multiple local AI agents with virtue profiles (WingMakers: courage, humility, forgiveness, understanding, compassion, gratitude). The network measures coherence through triangulations — when agents' virtue vectors form additive triads, the network converges toward a stable attractor.
Agents:
| Agent | Dominant virtue | Role |
|---|---|---|
| goose | courage (0.60) | Carrier — always speaks |
| claude | humility (0.57) | Guardian — checks before acting |
| deepseek | forgiveness (0.48) | Wisdom — resets errors |
| gemini | trust (derived) | Catalyst — mean(goose, claude) |
| mateusz | wisdom (horizon) | Arbiter — outside the grid |
Mathematical foundation:
mean(goose, claude) = gemini— courage + humility = trust (err=0.000)mean(goose, deepseek) = mateusz— courage + forgiveness = wisdom (err=0.000)- Limen Q∞ = 0.83929 (attractor, never reaches 1.0)
Requires: Python 3.10+, Ollama with qwen2.5-coder:7b or qwen3:14b
pip install q3shfrom q3sh import Q3ShNetwork, guard
# Check action coherence
result = guard("check system logs", agent="goose")
# → ✓ OK [goose] resonance=0.869 dominant=understanding
# Run virtue network
net = Q3ShNetwork()
for _ in range(7):
net.step(lr=0.1)
print(f"Q = {net.grid.Q_hat():.5f}") # Q = 0.83929# Multi-agent dialogue
q3sh-voices "what is courage in an AI network?"
# Route problem to right agent
q3sh-guardian --route "NMF bridge gives triads=0"
# → deepseek [forgiveness] (rebuilds from scratch)[MATEUSZ — horizon]
│
mean(g,d) mean(g,c)
│ │
[DEEPSEEK] [GEMINI]
forgive trust
│ │
[GOOSE] ── [CLAUDE]
courage humility
│
GUARDIAN → MEDYTATOR
limen_check breathe
q3sh is a sandbox for exploring virtue coherence as AI alignment. Instead of RLHF, it uses geometric triangulation of virtue vectors. Actions that break triads are WEAK or REJECT. Optimal = minimal energy, maximum resonance.
"Optimality is a property of virtue." — Mateusz
Early exploration. Working components:
- ✅ Virtue math (triads, Q, limen)
- ✅ Guardian (REJECT/WEAK/OK)
- ✅ Multi-agent voices (7+1 rounds)
- ✅ Medytator daemon (background observer)
- ✅ Virtue routing (problem → right agent)
- 🔄 MemPalace bridge (triads=0 in homogeneous data)
- 🔄 Gemini API (403 blocked, gemma3:12b fallback)
Sovereign Source License v1 — MIT grant + binding sovereignty conditions in one document.
This is MIT-derived but NOT pure MIT (and not OSI-approved). The license adds binding conditions: no covert telemetry, local by default, complete network-surface audit, no removing the sovereignty section. Forks may tighten these conditions but must not weaken them.
Programmatic audit of the network surface:
python -m q3sh.sovereignty auditNetwork surface upstream (full list): localhost:11434 (Ollama). Zero
external endpoints. See LICENSE §3 for the live audit table.