01 Writing

Writing

Notes on evaluating systems built around language models. Each post is about something I actually built, with the numbers, what broke, and what the fix cost.

Nothing published here yet. Two posts are in progress.

1. In progress

   The eval gate in Loop, and the one check it still misses

   How prompts get pinned by hash, how CI re-grades recorded model outputs offline, and the retrieval floors that are pinned in the repo but still checked by a command I run myself.

2. In progress

   The MIT-BIH split that learns the patient

   The same CNN scores 98.36% one way and 83.88% the honest way. What the beat-wise split leaks, and how the split search keeps all six classes in the test set anyway.

Back to the rest of the site