A simulation study positioning a 24-byte, estimation-free difficulty controller (LearningGate)
against Computerized Adaptive Testing, staircase rules, matched stored-weight twins, and frontier Large Language Models.
Experience the authentic WERR tripod Mandelbrot kernel dynamically adjusting step sizes in your browser.
How the 24-byte fractal-seeded controller compares with standard psychometric testing, staircase rules, and frontier AI.
| Dimension | Standard CAT | Fixed Staircase (= Elo) | PEST-type Rule | LearningGate (Fractal) | Frontier LLM (API) |
|---|---|---|---|---|---|
| Underlying Model | θ ~ N(0,1) Rasch | None | None | None (Estimation-Free) | In-Context Heuristic |
| State Footprint | 121-point posterior | 1 float | 3 floats | 24-byte seed + 6 floats | Multi-GB VRAM |
| Decision Latency | 0.008 ms (8 μs) | < 0.003 ms | < 0.002 ms | 0.3 - 0.4 ms | 67 - 788 ms |
| Stationary Tracking | 0.257 logits (Best) | 0.356 logits | 0.412 logits | 0.419 logits | 0.784 logits |
| Fast Learning Tracking | 0.755 logits lag | 0.401 logits | 0.465 logits | 0.421 logits (Strongest) | Lagged |
| Recovery After Jump | 56.0 trials | 14.8 trials | 11.7 trials | 10.4 trials (Fastest) | Unstable |