Deterministic inference and evidence

Portable arithmetic

1. Feature values are already fixed-point integers.
2. Each neuron sums integer weight × input products.
3. The sum is divided by the configured scale, truncating toward zero for positive and negative values.
4. The integer bias is added.
5. ReLU, when selected, replaces negative results with zero.
6. The result saturates to the configured positive or negative value limit.

No random state, locale, time, filesystem order, floating-point instruction, or network resource participates.

Evaluation evidence

The canonical evidence record contains schema and version, attachment id and fingerprint, bound Monk fingerprint, model revision and mode, ordered feature values, ordered output scores and thresholds, selected label, authority: advisory, and can_execute: false. Its SHA-256 is the evaluation fingerprint.

Cross-runtime parity

The Python and Pure-Perl implementations use the same ordering, signed truncation, saturation, JSON member ordering, UTF-8 encoding, and SHA-256 contract. The proof suite compares both evaluation fingerprints and both proposed attachment fingerprints.