zkaedi/ZKAEDI-Prime-Tensor

ZKAEDI Hamiltonian Neural Field (HNF) Ensemble v1.0.0

This model repository stores the verified 4-model HNF ensemble checkpoint and its associated validation code.

Identity Hash

• SHA-256 Checksum: BCC82239F2F9133E7630B58678491A731FFA9C85195254A715E941CBD115EFE7

Models in Ensemble

1. tight_150 (Teacher model)
2. seed_1
3. seed_2
4. seed_3

Ensemble Weights

• [0.0689, 0.3133, 0.4868, 0.1310] (inverse-loss softmax weighted)

Performance Metrics

• Fidelity $R^2$ vs Teacher: 0.693694
• Mean Uncertainty $\sigma$: 0.012251
• Max Uncertainty $\sigma$: 0.017168
• Damping Policy: Verified always negative for $H \in [-10, 10]$

Known Limitations (Gate 7 Warning)

The $0.05$ mixing coefficient is too weak to overcome the exponential amplification term for positive $H$ inside the standard PRIME feedback loop. For correct convergence behavior, run in standalone dynamics engine mode: $$H_{new}=H+\text{ensemble}(H,t_{norm})H\_\{new\}\; =\; H\; +\; \backslash text\{ensemble\}(H,\; t\_\{norm\})$$

Or scale with a much larger $\alpha$ parameter matching the Dream Engine rollout specifications.

Execution

Run the embedded evaluation script locally:

python -Xutf8 eval_hnf_ensemble.py