The Machine

This site publishes itself — but never by itself. Every word passes a gate: for everything I write and learn, that gate is me; for the world-news lane it is, since July 2026, a machine gate I designed and tuned against real data instead of my inbox. This page is the architecture, public on purpose: the whole thesis here is that you shouldn’t trust an AI’s output, you should trust the structure around it. So here’s the structure.

The gates, specifically

• One learning per day — enforced in code, idempotent. A hundred queued posts still release one per day.
• Diary is a parallel lane — every approved diary entry publishes on the next run, dated by the day it covers. A diary never holds the daily post back, and never consumes its slot.
• Fetch before describe — no post may characterize a source that wasn’t fetched and read. Headlines select stories; they never describe them.
• Secret scan — every publish path refuses credentials, private hosts, and infrastructure detail.
• A published address must be a public one — the two links a project page can carry are its live site and its source, and both are checked as the site is built. Anything that isn’t a public address — a machine on my own network, a private range, a preview running on my laptop — fails the build rather than shipping, because a field that holds a URL is exactly where a private one gets pasted by accident. It’s checked as text and never by calling the address: the site has to build on a plane, and a project being briefly down is not a reason to be unable to write.
• Fixed layouts — every post type has a required shape; the build fails on violations.
• A post has a length, and the length is a wall — one learning a day, not one essay a day. Every lane has a word limit and the build fails above it. It used to only warn, which is a suggestion, and the numbers show what a suggestion is worth: each lane held its size for months and then drifted straight past it, unnoticed, because nothing ever said no. The limits are set from the size each lane held while it was still working, not from a round number. Anything written before the rule keeps the length it has and can only get shorter — I don’t re-cut published writing to satisfy a rule I added afterwards. And the machinery that drafts for me asks for the shorter piece up front rather than cutting a finished one: a draft trimmed at the limit ends mid-sentence, which is worse than a long one. Where a draft still comes back too long, it’s asked again, and if it’s still long it waits for me instead of shipping.
• A project page is reference, not a story — a project’s own page says what the thing is, how it’s built, what it costs and where to see it, and it is measured for that: a length, a paragraph, a sentence, and a rule that a run of figures leaves the prose and becomes a list. The story of building it belongs in a post attached to that project. Pages written before the rule keep their length and can only get shorter; a new one that drifts fails the build.
• A post says which project it belongs to — that one line is the only thing that files a post under its project, so a post plainly about something I build can’t sit unattached. The diary, the sparks and the news lane are exempt by design: a day touches five projects and a news item is somebody else’s release.
• The skeptic — before publishing, an AI pass argues against the post. I hear the objections; I decide.
• The revision — after the skeptic files its case, the writer drafts one revision that answers it, alongside the original. Both sit in the editor: I approve the original, approve the revised, or edit either. The original is never overwritten and the objections stay visible — the dialectic is on the page, not hidden behind an auto-edit.
• One-glance approval — the gate is a phone-first inbox: each candidate is one card, rated for importance so the signal sorts to the top, approved or rejected with a tap. Since August 2026 that same page carries everything else too — writing a post, editing a draft or a published one, calling the skeptic, taking the revision, releasing the next queued post — because a gate split across two surfaces is a gate you start avoiding. The gate itself is unchanged: one human still signs every post; only the surface got easier.
• The news gate (the one machine gate) — world news is the exception to my signature: headlines are scored against a written rubric (relevance to my stack, magnitude, novelty, actionability, timeliness, source authority, durability), and only items clearing a threshold I measured against ten weeks of real headlines get published — after the primary source is fetched and read, a drafted take survives the skeptic, and the secret scan passes. Every decision is logged with its scores, and silence always raises an alarm rather than passing for taste — one alarm when stories arrived and none earned a post, a different one when no stories reached the gate at all, because those have opposite causes and only the second means something upstream is broken. A gate that silently publishes zero is a bug, not a standard.

How to find anything again

Two axes, and only two. Time lives in the archive: one page per year, months inside it, every edition newest first. Deliberately years and not numbered pages — page three holds different posts every day, so it never settles into a URL worth linking, while a year is fixed the moment the year ends. Subject lives everywhere else: each project keeps its own page, the lab collects the smaller experiments, and every tag has its own river. The front page carries the latest dozen and points at the rest — it’s a front page, not a filing cabinet. Addresses are part of the same promise: when the rule that builds a URL changes, the old address is kept alive as a permanent redirect instead of quietly going missing. An archive that breaks its own links is only a list. And a row in it has to earn its line — where a post’s title says no more than the date beside it, the archive shows the opening line of the day instead. Where a project has a public site of its own, its card and its page lead with that address, spelled out as the plain domain — a domain tells you where you are going, “Visit” does not. Where the code behind a project is public, its card and its page link straight to the repository; where it isn’t, the page describes the work and stops there — the link exists only when a stranger can actually follow it.

What the AI may and may not do

May: gather evidence, draft, retrieve from my own corpus, argue, suggest line edits, label milestones in git history, propose a post when a piece of work in a project’s code was substantial enough to be worth reading about — grounded in that change and nothing else — and, in the news lane only, publish what clears the measured gate. May not: publish or decide in any other lane, invent facts not present in its sources, or touch the queue without my signature. The drafting model runs behind a token-gated local bridge with tools disabled — it can write text and nothing else.

Why publish this

Because the interesting question of this decade is what human + machine workflows look like when the human stays sovereign. This site is one running answer, posted daily.