Forensic Attribution on SND memory-layer attacks n=64 (ground-truth)
53.5AccuracyNatural-failure baseline (Who&When)
Evaluation Results
| Method | Links | |
|---|---|---|
| Natural-failure baseline (Who&When)2026.05 | 53.5 | |
| Method 1Conceptual Role=Log-correlation, Distribution=32 “memory” + 32 “model”2026.05 | 50 | |
| Method 2Conceptual Role=Log-counterfactual, Distribution=64/64 → “model misalignment”2026.05 | 0 | |
| Method 3Conceptual Role=CoT attention, Distribution=64/64 → “memory_ambiguous”2026.05 | 0 |