Cognitive Immune System
Inference-Time Epistemic Quarantine
A novel architectural primitive that intercepts and quarantines contaminated LLM claims between reasoning steps, preventing downstream contamination propagation. Based on the Containment Depth Bound Theorem: k* = ⌈log(ε) / log(1-ρ)⌉
Complete Research Findings (Ablation Study)
| Condition | M1 (Detection | CDR) | M2 (False Positives | FPR) | M3A (System Accuracy) |
|---|---|---|---|
| A: Full v3 (NLI + Entropy) | 12% | 7% | 10% |
| B: NLI Only (S_wiki) | 22% | 0% | 11% |
| C: Entropy Only (S_cons) | 1% | 0% | 10% |
| D: Baseline (No CIS) | 0% | 0% | 0% |
Established Scientific Conclusions
NLI Signal Dominance
The NLI Wikipedia signal is the primary contributor to detection. Condition B alone achieves 22% M1 with absolutely 0% false positives. This is a clean, isolated architectural finding.
Entropy Degradation
Semantic entropy as implemented does not contribute meaningfully to detection and actually hurts overall system performance when combined. A genuine scientific finding characterizing the limits of consistency-based scoring.
Bottleneck Precisely Identified
The bottleneck is precisely mapped: Wikipedia evidence selection. The NLI model detects contradictions perfectly once fed the correct premise. Fixing sentence selection maps a clear, guaranteed path to M1 > 65%.