Interaction integrity monitoring on LLM interaction perturbations
100Detection CoverageIDT (P)
Evaluation results
| Method | Links | |
|---|---|---|
| IDT (P)What is computed=Token-frequency distributions → entropy / mutual information terms → P, Auxiliary model required=None, Operation per turn=Count tokens; update information ratios, Context handling=No context reprocessing; uses token distributions only, Scaling with interaction length=Low; depends on token counting/aggregation, External dependency=None, Primary failure detected=Informational uncoupling2026.03 | 100 | |
| Cosine similarityWhat is computed=Text embeddings → cosine similarity, Auxiliary model required=Embedding model, Operation per turn=Forward pass through encoder, Context handling=Requires text encoding, Scaling with interaction length=Depends on encoder input length, External dependency=Local or external encoder, Primary failure detected=Representational similarity loss2026.03 | 100 | |
| LLM-as-a-judgeWhat is computed=Prompt/response (+ rubric) → evaluator score, Auxiliary model required=Evaluator LLM, Operation per turn=Full inference call, Context handling=Requires full prompt/context submission, Scaling with interaction length=High; depends on evaluator prompt length, External dependency=Local or API-based evaluator, Primary failure detected=Evaluator-scored quality/rubric mismatch2026.03 | 100 |