Read everything. Keep almost nothing.
What the machine reads, how it ranks and cuts, and how to check it against its own ledger.
There is too much to read, and most of it isn’t worth reading. Eclecta reads it anyway. It watches the places technology, AI, and the sciences break first (the labs, the front pages, the long tail of practitioner blogs) and files the few stories that matter, each with a plain account of why. About one cluster in 122 survives to become a pick. The cuts are the product.
How it works
Everything starts wide. The pipeline pulls from about 3,200 verified sources: RSS feeds plus Hacker News, Lobsters, Reddit, arXiv, Mastodon trends, Google News, Wikipedia current events, and GDELT. Items are canonicalized and clustered, so one story reported on a dozen surfaces counts once. A deterministic score then ranks each cluster on consensus, engagement, reputation, recency, and topic fit. No model has touched anything yet.
Then three do, in rising order of cost. Claude Sonnet 5 triages the ranked clusters down to finalists. Claude Sonnet 5 reads each finalist in full, pulls the load-bearing details, and writes the one line that says whether to care. Claude Opus 4.8 writes the editions, under a style guide with no appeals process. Those names come from the pipeline’s own configuration; when a stage changes models, this paragraph changes with it.
The editions run on a fixed clock: the daily brief on weekday mornings, the weekly digest on Fridays, then the monthly review, the quarterly report, and the year. Each cadence answers its own question. The daily says what happened; the weekly names the through-lines; the retrospectives weigh what proved to matter.
The pipeline writes what it publishes into this site’s repository and pushes; GitHub Actions builds and deploys the result. The site is static. If the machine stops, the site stands; it stops getting newer.
What earns a slot
Every edition is written for one reader: a senior engineer who wants primary sources, technical depth, and no theater. A story earns a slot on consequence (it changes what a practitioner builds, ships, secures, buys, or believes), on novelty, on evidence, or on durability, and it ranks by how many of those it clears and how hard. By policy, a day narrows to about 150 finalists; an edition keeps a handful.
The cuts are by standing rule: marketing dressed as news, point-release churn, engagement bait, vendor benchmarks without methodology, and thin rewrites of someone else’s reporting. When unsure, cut.
What survives is handled by rule. Claims stay attributed claims until independently checked; preprints are labeled; self-reported benchmarks are flagged every time. Where credible sources disagree, both run, with names attached. The primary source is linked first, always. When the original sits behind a paywall, a free read is linked beside it.
Why trust it
A publication written by machine earns trust the way an instrument does: by being checkable. Do not trust it. Check it. The coverage page shows what gets read, which surfaces are loudest, and the pipeline’s numbers, down to which model ran each stage; the sources page names every feed on the roll. The figures here (the source count, the survival ratio, the model behind each stage) are read from that same ledger, regenerated at each publish, most recently on August 17, 2026.
Open source
The machine that writes the ledger is itself open. Every part of it (the fetchers, the deterministic score, the prompts, the rubric that cuts, and the code that renders this page) lives in one public repository under the MIT license. Read it, run it, fork it, or lift a piece for your own. The claims on this page (no tracking, a fixed clock, these models at these stages) are not requests for trust; they are lines you can go and find. A publication you cannot inspect is just a louder opinion. This one you can inspect, and it stays that way.
What it costs
Nothing, in money or in data. There is no account, no tracking, and no analytics: the site does not measure attention, so it has nothing to sell. Reading preferences (theme, type size, density, muted sections, the signals panel) are stored in this browser’s localStorage and nowhere else; nothing about a visit leaves the device. Close the tab and the relationship ends.
Subscribe
Eclecta is built to live in a feed reader. Point one at the full feed, take the digests alone, or pick a single cadence or section from the feed directory. Picks arrive as they’re curated; editions arrive on their clock. Every edition published stays in the archive.
Colophon
Schibsted Grotesk sets the headlines, Source Serif 4 the body, IBM Plex Mono the furniture; all three are self-hosted. No images anywhere, only type, spacing, and hairline rules. Astro 5 builds the site; GitHub Pages serves it. The site and the pipeline behind it are open source, MIT licensed.
Eclecta is built and run by one person, Illya Starikov. Corrections, broken links, and bad calls go through the contact page. The words are machine-written; the responsibility is not.