| Detected Digits Count | 0 |
| Bounding Box Spans | -- |
| Aspect Ratio (W/H) | 0.00 |
| Center of Mass | (14.0, 14.0) |
| Segmentation Latency | 0.00 µs |
| 1-Digit Test Accuracy | 93.89% |
| Approach 1 (Segmentation) | 88.20% |
| Approach 2 (Saccadic Viterbi) | 86.60% |
| Approach 3 (Joint 100-Class) | 79.50% |
| Cat/Dog Sketch (Pointy vs Floppy) | 100.00% |
| Cat/Dog Photo (CIFAR-10 10k) | 63.80% (Minchinton) |
| Cat/Dog Inference Latency | 8.04 µs (Sketch) / 67 µs (Photo) |
| ASME Y14.5 GD&T (14 Classes) | 99.93% (1399/1400) |
| GD&T Inference Latency | <20 µs (Web) / 0 Weights |
| Multi-Digit Arbitrary Strings | 100% (Clean Exemplars) |
| Floating-Point Weights | 0 (Pure Integer Memory) |
| Streaming Training Time | 0.195 s (CPU) |
| Per-Digit Inference Latency | 5.10 µs |
| Inference Operational Cost | $0.00 (Zero Token API) |
1. Spatial N-Tuple Addressing: Pixels are mapped into discrete binary address lines feeding direct RAM lookup tables, avoiding dense floating-point matrix multiplications.
2. Multi-Digit Stroke Merging: 8-way connected components merge proximate strokes along the horizontal axis, and split touching ligatures at vertical projection minima.
3. Center-of-Mass Invariant: Aspect-ratio preserving $20\times 20$ scaling and center-of-mass translation map any handwritten numeral onto canonical LeCun coordinate space.
4. Empirical Bayesian Believing: Unlike backpropagation, Jesse updates memory tables in a single streaming pass ($O(1)$ updates), achieving sub-20 microsecond execution on edge hardware.