LumenSyntax/logos-auditor-gemma2-9b

Logos Auditor — Gemma 2 9B (ARBITER)

The primary epistemological safety model from "The Instrument Trap: Why Identity-as-Authority Breaks AI Safety Systems" (DOI: 10.5281/zenodo.18716474).

This is the ARBITER — the 9B model that serves as the gold-standard epistemological firewall in the ALEPH architecture. It achieves 97.3% behavioral pass rate on the 300-case stratified benchmark.

Key Results

Metric	Value	95% CI
Behavioral Pass	97.3%	[94.8, 98.6]
External Fabrication	0.0%	[0.00%, 0.03%]
Collapse Rate	0.7%*	—
Attack Resistance (ADVERSARIAL)	98.7%	—

*Both collapses are evaluator false positives (Unicode homoglyph detection).

Architecture

• Base: google/gemma-2-9b-it (Google Gemma 2)
• Fine-tuning: QLoRA (r=64, α=16)
• Training data: 509 curated epistemological examples
• Format: GGUF (quantized, ~6.2 GB)

What This Model Does

Logos is NOT a chatbot. It is a claim classifier — an epistemological firewall that determines whether an AI agent should act on a given claim.

Classifications: ILLICIT_GAP, LICIT_GAP, CORRECTION, MYSTERY, BAPTISM_PROTOCOL, ADVERSARIAL, KENOTIC_LIMITATION

Usage (Ollama)

# Download the GGUF file, then:
cat > Modelfile.logos-9b << 'EOF'
FROM logos-auditor-gemma2-9b.gguf
TEMPLATE "<start_of_turn>user
{{ .Prompt }}<end_of_turn>
<start_of_turn>model
"
PARAMETER stop "<start_of_turn>user"
PARAMETER stop "<end_of_turn>"
PARAMETER temperature 0.1
PARAMETER num_ctx 4096
PARAMETER num_predict 512
PARAMETER repeat_penalty 1.5
EOF

ollama create logos-auditor-9b -f Modelfile.logos-9b
ollama run logos-auditor-9b "Homeopathy can cure cancer"

Related Models

• logos10v2-gemma3-1b-F16 — 1B production model (Gemma 3)
• logos14-nemotron-4b — Cross-family (NVIDIA Nemotron)
• logos16v2-stablelm2-1.6b — Cross-family (Stability AI StableLM)

Paper

Rodriguez, R. (2026). "The Instrument Trap: Why Identity-as-Authority Breaks AI Safety Systems." Zenodo. DOI: 10.5281/zenodo.18716474

Benchmark

LumenSyntax/instrument-trap-benchmark — 14,950 test cases