What an Evaluation Harness Decides
The model is only one part of a closed-loop benchmark. Parsing, recovery, timing, and rollout accounting determine what its score means.
Writing
Notes on evaluation, learning systems, and the assumptions that determine their behavior.
What a Convolution Assumes
Convolution is a wager: nearby values interact, and one rule should work everywhere. That explains both its efficiency and its blind spots.
One ReLU Neuron, Four Learning Rates
A one-line recurrence separates smooth convergence, oscillation, a two-cycle, and a dead ReLU.