Classical Energies Learn to Reason 1,000x Faster than Looped Models
Paper β’ 2611.17309 β’ Published
Pure-numpy CRAM engine (audit of arXiv:2611.17309) with energy certificates on every answer. Source and full audit trail: https://github.com/szl-holdings/szl-cram
model100.npz β 100-epoch pseudolikelihood fit, 9x9 (W, B, threshold). 131,949 nonzero parameters, fitted threshold -208.25.grid2_models.npz β grid-2 refit (9x9) used for the cfgA/cfgB evaluations.| Configuration | Exact | Time |
|---|---|---|
| Learned energy, paper-style search | 0/150 | 234 s |
| Learned energy, MFA v2 (ours) | 1/150 (0.67%) | 695 s |
| Hand-set energy (1 weight, 0 trained params), MFA v2 | 13/150 (8.67%) | 1032 s |
| Hand-set energy, focused Gibbs | 4/150 (2.67%) | 98 s |
Key finding: positives-only pseudolikelihood fits create spurious deep minima at solution depth (interp9: min off-diagonal cost -0.5506, i.e. decoy REWARD entries). The hand-set one-weight energy beats the learned fit at the extreme end; we publish both.
Every answer ships {answer, E_final, threshold, margin, sweeps, valid}. The threshold is
fitted from training-solution energies; margin > 0 means "at least as good as a training
solution, probabilistically". Machine-checkable, deny-by-default compatible.
SZL Holdings / a11oy. Apache-2.0.