Everything you see is a projection of something you don’t. A person’s behavior is a projection of emotional systems they can barely articulate. A file on a screen is a projection of canonical state stored somewhere else. A company’s products are a projection of its org chart. The surface is never the system.

The Projection is a running read on three things: how AI inference economics actually work underneath the headline, what governed infrastructure gets right that ad-hoc infrastructure doesn’t, and how to tell a widely repeated claim from a well-supported one. It’s written by a clinical psychologist who also builds software — the through-line across all three is the same one.

Threads here are published, not generated on the spot — each one is a narrative tracked over time, with dated evidence and a link back to its source, curated daily under a fixed rubric: signal over noise, sourced claims only, no framing spin. The editorial judgment happens when something starts being tracked, not as a separate approve-before-it-goes- live gate — once a thread exists, it publishes automatically. Nothing here is un-curated; it’s just not hand-gated a second time on the way out.

How it’s built. This site is a thin, static export. All the real work — sweeping public sources, curating each day, tracking threads over time — happens in kestrel, a private tool that never publishes anything itself. What you’re reading is generated from kestrel’s tracked state and pushed here: Hugo renders it, Cloudflare serves it, nothing runs server-side. The site’s own code is public and MIT-licensed if you want to see exactly how a thread becomes a page.

What’s at the top of every page. Each page opens with a readout in three parts: Breaking (anything from today), News (the last seven days), and a short Summary. The first two are assembled mechanically from the same dated item record the threads themselves are built from — no model writes them, so they can’t drift from what the sources actually say. The summary is different: it’s written by a language model against that page’s tracked material, and regenerated only when the evidence underneath it changes. Treat it as orientation, not citation — the dated bullets and their source links are the record.

Above all of that, rarely, a flash — a single banner carrying something that would lead a general news front page whether or not it touches AI, global capital, or mental health. The bar is deliberately high: most days there is none, and if you start seeing them often, that’s a bug rather than a busy week. A flash can be dismissed; it returns on reload.

How I use it. This started as a private morning read — three lenses I actually track, so I’m not reconstructing context on a news cycle from scratch every day. What’s here is that same read, minus whatever’s genuinely just mine to know. If you want to dig into a thread rather than just read it, every thread page has a button that packages its whole tracked timeline for pasting into whatever AI chat you already use — that’s how I actually interrogate one when something looks off.

Questions, corrections, or something worth tracking that isn’t here yet: open an issue on the site’s repo.