WEBVTT

00:00:00.000 --> 00:00:08.667
Can multiplying by one change an answer? In this MLX CPU test, an optional unit weight changes the output precision.

00:00:08.667 --> 00:00:20.167
For a zero input, normalization contributes zero. A finite bias of seventy thousand should remain seventy thousand. Without weight, the original returns infinity.

00:00:20.167 --> 00:00:28.733
Adding a unit weight preserves that bias. The output type branch ignores the bias type when weight is absent, narrowing it before addition.

00:00:28.733 --> 00:00:38.900
The bias gradient can lose the same finite values. The candidate includes the bias in type promotion, preserving this tested forward and backward result.

00:00:38.900 --> 00:00:50.813
A clean CPU build reproduces one cause across three hundred eighty five assertions: one hundred six failures before, zero after. Sixteen controls remain unchanged.

00:00:50.813 --> 00:01:01.520
This is pinned native C plus plus evidence. GPU and installed wheel behavior remain unverified. Code, logs and the candidate patch accompany the report.