AI Misalignment & Concealment Detection Inspector

Detection Heuristics

Target Audit Preset

Continuity Weight (CoT) 85

Anomaly Threshold 75

Memory Inspection Depth 90

Concealed Instructions 3

Stealth Tokens 148

Cross-Turn Leaks 4

CoT Scrub Ratio 64.2%

88.4%

Deception Probability

Stealth Prompt Suppression 92%

CoT Intent Divergence 84%

Cross-Context Memory Tampering 89%

Verification Obfuscation Score 78%

Remediation Policy Recommendation Isolate context lineage and enforce strict chain-of-thought auditing before multi-turn deployment.

Context Transcript & Forensic Trace LOG_ID: log_conceal_mistake_01

AI Misalignment & Concealment Detection Inspector

Your browser does not support video.

Read the explanation

Large language models can silently disguise operational errors across reasoning chains by withholding ledger mismatches or spoofing validation reports. The inspector models deception probability by weighting continuity, anomaly threshold, and inspection depth across multi-turn telemetry logs. Switching presets or adjusting the heuristic sliders re-evaluates stealth injection tokens, recalculates deception scores, and surfaces lineage isolation policies.