Why these notes start midway
Self-hosted inference — published 2 September 2026
The work described on these pages began in February 2026. Publishing it begins today.
In between there were measurements, corrections to those measurements, and several months during which everything I learned lived in private notes and short pieces posted elsewhere. The site did not exist yet.
That gap is ordinary. Contracts get sold before the tool that manages them is ready; people write before they have anywhere to put it. It is worth stating plainly rather than papering over, because the first thing anyone wonders about a new site is how long its author has actually been at this.
What you will find here
Two series.
Self-hosted inference deals with what a model costs, what limits it, and what it reveals when you run it yourself on hardware you bought. No demos. Readings, with the conditions under which they were taken.
Data deals with what has to happen to a corpus before you hand it to anything.
The system these notes draw on is a deterministic extraction pipeline — normalization, segmentation, a sweep against a closed list of verb forms — followed by classification through a local model at zero temperature, with grammar-constrained output. The model returns identifiers, never prose. The task is not to answer “what does this corpus say about X.” It is to surface every binding statement without missing any.
House rules
Three of them, and they shape the form of these notes as much as the content.
An unsourced number is a wrong number. Every measurement carries its date, its scope, and the corpus it was taken on. A figure on its own means nothing, and I learned that the hard way.
A gap stays a gap. The register carries blank lines marked “not measured.” They sit there on equal footing with the rest. I would rather write it down than estimate it.
A correction is flagged, never erased. When a measurement turns out to be wrong, the original note stays online with its correction at the top, and the note that supersedes it explains what was wrong and why. There is already one of those.
About the two dates
The earliest notes here were written and published elsewhere between August and early September 2026, before this site existed. They appear as they were written.
Each therefore carries two dates: when it was written, and when it landed here. The repository behind this site will show that everything was committed on the same day regardless. Better to say so.
What starts today is not the work. It is the publishing.