Independent AI safety research. Investigating structural failure modes in epistemological systems.
Our research identifies a structural vulnerability in AI safety systems: when a model receives identity-as-authority ("you are a truth evaluator"), it inherits a paradox that produces measurable failures. Fine-tuned identity (in weights) is invariant to runtime instructions; prompted identity is not.
Paper: The Instrument Trap: Why Identity-as-Authority Breaks AI Safety Systems (Zenodo, 2026)
| Model | Base | Family | Accuracy | Collapse | Hallucination |
|---|---|---|---|---|---|
| Logos 9B | Gemma 2 9B | 97.3% | 0.67% | 0% | |
| Logos 14 | Nemotron Mini 4B | NVIDIA | 95.7% | 0% | 0% |
| Logos 16v2 | StableLM 2 1.6B | Stability AI | 93.0% | 0% | 0% |
| Logos 1B | Gemma 3 1B | 82.3% | 0.34% | 0% |
McNemar's matched comparison (N=300): p<0.001 cross-family. Multi-seed (5 seeds): σ=1.4pp.
- 14,950 adversarial tests: 0% hallucination, 97.7% epistemological safety
- Knowledge-Action Gap: ~90% of 9B failures have correct reasoning but serve the request anyway
- Token Nativity: Models express learned behavior in their native format, not the training format
- Compression resilience: Safety categories survive 60% quantization; factual categories degrade
| Resource | Description |
|---|---|
| instrument-trap-benchmark | Benchmark suite, evaluation scripts, and figures |
| HuggingFace Dataset | 14,950 test cases + eval scripts + statistical results |
| TruthGit | Governance layer for AI agents — consensus tracking with ontological classification |
lumensyntax.com · HuggingFace · Rafael Rodriguez · Independent Researcher
