Precision & Dtype Notes (v0.4 SCM)

ZeroProofML v0.4 is SCM-first and represents singularities via an explicit ⊥/mask rather than IEEE NaN/Inf arithmetic. Precision still matters, especially for detecting “near-zero” denominators and for stable optimisation.

  • Prefer float64 (torch.float64) for rational heads (SCMRationalLayer) when your workload is near-singular (robotics IK, stiff physics).
  • Keep denominator singular detection explicit via the layer’s singular_epsilon (see zeroproof/layers/scm_rational.py).

Example:

import torch
torch.set_default_dtype(torch.float64)

What is (and isn’t) enforced

  • There is no global float64 enforcement layer in the v0.4 core.
  • The Torch examples set dtype explicitly; your application should do the same.
  • Mixed precision is supported in the trainer (zeroproof/training/trainer.py) via AMP; for non-CUDA devices the default AMP dtype is adjusted away from float16.

Practical guidance

  • If you see unstable bottom_mask behaviour (too many / too few singular flags), tune:
  • SCMRationalLayer.singular_epsilon (forward detection)
  • trainer tau_train_min/tau_train_max (projective denominator thresholding)
  • When exporting to IEEE floats, ⊥ maps to NaN (zeroproof/utils/ieee_bridge.py) so downstream metrics tooling can treat it as invalid.