Mini GPT
A character-level GPT written from scratch in plain Java with no ML library at all — bigrams, then backprop by hand, then a small transformer.
Write-ups with measured numbers.
Numbers taken from records rather than recollection, and a black box taken apart on screen. Every input is shown, so a figure is something to check rather than to trust.
A character-level GPT written from scratch in plain Java with no ML library at all — bigrams, then backprop by hand, then a small transformer.
What a Claude Code task really costs, and how far each plan carries you — every coefficient measured from real session transcripts.
Where the hours actually went on an AI-assisted project, attributed from session records and git history rather than recalled.
How to run a project where Claude Code writes the code and a person decides what gets built, with every rule resting on graded evidence.
Rule pages render a bold normative sentence holding an italic term as strong with em inside, and a test fails on any rule page that prints two asterisks.
Every page's header carries the "part of alegauss" chip to alegauss.github.io and a GitHub button to the repository, and a test fails on a header without either.
markdownlint now ignores docs/DEFERRED.md beside the other four governed files, and its gate passes with the file as roadkeep wrote it.
The guide and spec pages take fonts, tokens and site.css at 2026.3.22, re-keyed to the mark's rose, with analytics behind consent; style.css keeps only this site's own blocks.
Every built page links icon.svg and carries og:image, the guide header shows the mark, and a stdlib-drawn og.png renders the mark and title.
F410 grades Shio's SH1141 at R1/S3: files cost 1.29x less and the agent writes 3.1x less; AP's open question and the guide say so, and corpus.md now lists a commit cited past a pin.