Optimization Guide (SCM)¶

ZeroProofML v0.6.0 focuses on SCM-native performance rather than transreal profiling hooks. This guide highlights the current levers for efficient training and inference.

1. Prefer vectorised SCM ops¶

Use the backend factories in zeroproofml.scm.ops to keep payloads and ⊥ masks aligned without Python loops. Example:

import numpy as np
from zeroproofml.scm.ops import scm_add_numpy, scm_div_numpy

values = np.array([1.0, 0.0, -2.0])
mask = np.array([False, True, False])
num, num_mask = scm_add_numpy(values, values, mask, mask)
den, den_mask = scm_div_numpy(num, np.ones_like(num), num_mask, np.zeros_like(num_mask))

Masks remain boolean, making it easy to accumulate coverage metrics.

2. Track coverage, not branches¶

SCM graphs avoid guard rails; instead, record how often the model predicts ⊥:

bottom_rate = float(bottom_mask.float().mean())
if bottom_rate > 0.1:
    print("Too many singular predictions; adjust loss weights or epsilon.")

3. Use gradient policies intentionally¶

GradientPolicy.PROJECT zeros gradients on ⊥ paths; CLAMP caps finite gradients. Choose per-layer:

from zeroproofml.autodiff.policies import GradientPolicy
from zeroproofml.layers import SCMRationalLayer

head = SCMRationalLayer(1, 1, gradient_policy=GradientPolicy.PROJECT)

4. Projective tuples during training¶

Keep rational heads in projective form during optimisation and decode with the ⊥ mask at evaluation. This preserves stability near poles without hiding coverage gaps.

5. Keep IEEE ingress explicit¶

Map NaN/Inf to ⊥ at the boundary using zeroproofml.utils.ieee_bridge. Avoid ad-hoc nan_to_num calls that would erase mask information.

6. Benchmark with real payloads¶

Scripts in perf/ exercise the SCM backends without any transreal tags. Start from perf/run_benchmarks.py and track ⊥ throughput alongside wall-clock time.