BENCH-005: Ternary vs Binary — Extended Multi-Dataset Validation

[gHashTag]

Objective

Validate H1 on two datasets (MNIST + CIFAR-10) across 5 number formats.

Datasets

Dataset	Images	Classes	Resolution	Channels	Complexity
MNIST	70k	10	28×28	1 (gray)	Low
CIFAR-10	60k	10	32×32	3 (RGB)	Medium

Formats Under Test

Format	Bits	Bytes/weight	Compression vs FP32
FP32	32	4.0	1× (baseline)
GF16	16	2.0	2×
FP16	16	2.0	2×
BF16	16	2.0	2×
Ternary	2	0.125	32×

Model Architecture

• MLP: 784→128→10 (MNIST) / 3072→256→128→10 (CIFAR-10)

Success Criteria

• GF16 gap vs FP32: ≤0.5% on both datasets
• Ternary gap vs FP32: ≤2% on MNIST, ≤5% on CIFAR-10

Implementation Steps

1. Create src/cifar10_loader.zig — parse CIFAR-10 binary batches
2. Extend src/bench_format_comparison.zig — multi-dataset benchmark
3. Run 5 formats × 2 datasets × 3 seeds = 30 runs
4. Generate results table + CSV export
5. Update docs/proposals/BENCH_005_RESULTS.md

---

φ² + 1/φ² = 3 | TRINITY