Differential Diagnosis for AI Failures
Distinguishing hallucinations from infrastructure stress. When AI fails, knowing why determines what to do. High CEG means the model is lying. Low CEG means the system is struggling.
Hallucinations and infrastructure stress produce opposite signatures. TCEG's patented analysis distinguishes between them — enabling the right response to the right problem.
The key insight: Different failures require different responses. TCEG's patented methodology distinguishes between them automatically — it doesn't just detect problems, it diagnoses them.
Independently validated across multiple domains with consistent results.
| Method | Detection | Speed | Compute | Explainability |
|---|---|---|---|---|
| TCEG | 0.87 AUC | Real-time | ~10% | High |
| Semantic Entropy | ~0.80 AUC | 5-10× slower | High | Medium |
| Static Thresholds | Variable | Fast | Low | High |
| ML Classifiers | Training-dependent | Medium | Medium | Low |
TCEG achieves competitive accuracy with dramatically lower computational overhead through single-pass temporal gradient analysis.
One framework, multiple domains: TCEG for temporal streams, P-CEG for static images.
Real-time hallucination detection for ChatGPT, Claude, GPT-4, and enterprise LLMs. Single-pass detection at 10× lower compute cost than multi-sample methods.
Predict drug sensitivity in hours vs 24-72hr conventional endpoints. Temporal entropy gradients detect cellular response trajectories in real-time screening.
Early detection of sepsis, heart failure decompensation, and transplant rejection. Trajectory analysis catches critical transitions before threshold crossings.
13× better zero-day detection. Filter false positives by entropy trajectory: true threats drift upward, noise spikes and falls. Reduce SOC alert fatigue.
Flash crash prediction and market manipulation detection. Differential diagnosis distinguishes algorithmic spoofing from genuine market stress.
P-CEG authenticates images via perturbation stability. Three-class discrimination: authentic photographs, AI-enhanced, or fully AI-generated deepfakes.
Complete IP protection: TCEG for temporal streams, P-CEG for static images. Filed January 2026.
Foundational TCEG methodology for hallucination detection via temporal confidence-entropy gradients. Real-time streaming AI analysis.
AI-assisted diagnostic systems with uncertainty quantification. Identifies "confident guessing" requiring human review.
Threat detection in SIEM systems and black swan events in financial markets. Entropy trajectory analysis.
Perturbation stability for static image authentication. Three-class: authentic, AI-enhanced, or AI-generated deepfakes.
Temporal entropy gradients predict drug sensitivity in vitro. Classification in hours vs 24-72hr conventional endpoints.
O(1) bounded-memory temporal gradient computation. Circular buffer architecture for real-time streaming anomaly detection across all modalities.
Six UK patents protect complementary frameworks. TCEG (temporal) for streaming AI, medical monitoring, cybersecurity, and cellular drug response. P-CEG (perturbation) for static deepfakes. Incremental Computation for O(1) bounded-memory processing. Complete coverage from LLM safety to pharmaceutical screening.
From AI safety to pharmaceutical screening in 15 days: same core framework, domain-specific calibration.
See the differential diagnosis in action: watch CEG distinguish hallucination from infrastructure stress in real-time.
Demo Launching SoonKevin Wharram
Independent Researcher
TCEG emerged from a cross-disciplinary observation: Trajectory matters more than absolute value. After 25 years in cybersecurity and 18 years with SIEM systems, I noticed that benign threats spike and fall, while real threats drift upward.
The differential diagnosis discovery came from asking the wrong question in the right way: "Can CEG detect infrastructure stress?" The answer was no, but the reason why revealed something more valuable. Infrastructure stress and hallucination produce opposite signatures, making CEG a diagnostic tool, not just a detector.