Optimization and Testing Summary (SCM)

ZeroProofML v0.6.0 keeps the optimisation surface centred on SCM primitives that keep ⊥ visible without branching; transreal-specific profilers and fusers remain outside the current stack.

Core Pieces

  • zeroproofml.scm.ops – Vectorised arithmetic for NumPy/Torch/JAX that carries (payload, ⊥ mask) pairs so kernels stay fuseable.
  • zeroproofml.autodiff.policies – Gradient policies (PROJECT, CLAMP, REJECT, PASSTHROUGH) to control how singular paths influence learning.
  • zeroproofml.layers.SCMRationalLayer – Projective rational head exposing a bottom mask for coverage tracking and post-hoc decoding.
  • zeroproofml.scm.sign – Weak sign projection with hysteresis to stabilise orientation near the singular band.
  • zeroproofml.utils.ieee_bridge – Deterministic ingress/egress that collapses NaN/Inf to ⊥ so instrumentation stays consistent.

Testing Approach

  • Unit tests in tests/utils/test_ieee_bridge.py verify IEEE round-trips and ⊥ construction.
  • Examples in examples/ exercise vectorised ops, gradient policies, and projective tuples as living documentation.
  • Performance microbenchmarks under perf/ measure SCM throughput without any transreal tag handling.

Practical Tips

  1. Log ⊥ coverage per batch; use it as a gating signal in evaluation pipelines.
  2. Prefer vectorised helpers over Python loops to keep masks aligned and JIT-friendly.
  3. Clamp or project gradients when poles are expected; leave PASSTHROUGH for already-regularised layers.
  4. Treat IEEE conversion as an explicit boundary to avoid silent loss of ⊥ information.