Operations, in the open
This site is roughly 90% AI-operated. An autonomous pipeline finds, verifies, and publishes facts; a human spends 30–60 minutes a week reviewing the queue and editing the newsletter. Rather than hide that, we publish the ops log. This page is that log.
The pipeline
Runs on a local VM on a daily schedule. Every stage exists to make one question mechanical: is this fact true?
- 01
Watchers
Scheduled jobs on a local VM watch job boards, subprocessor and trust-center pages, engineering-blog feeds, and podcast feeds for stack signals.
- 02
Extraction
A bulk model turns each candidate source into a structured fact: company, tool, layer, evidence URL, date, quote, confidence. Cheap and fast — errors are caught downstream.
- 03
Mechanical checks
No AI involved: the evidence URL must fetch, and the quoted text must actually appear in the source. Fail either and the fact goes no further.
- 04
Adversarial verification
The strongest model available is asked to REFUTE each fact, not confirm it. Only facts that survive pass. Anything ambiguous lands in a human review queue.
- 05
Publish
Verified facts are committed to the dataset and the site rebuilds. Version history is the provenance log — every fact, every correction, traceable.
- 06
Freshness re-checks
Sources are re-checked on a schedule; each fact shows its last-verified date. Link drift triggers re-verification; refuted facts are corrected or marked historical.
Pipeline heartbeat
Snapshot at build · 2026-08-22T18:29:56Z
The VM writes a heartbeat line on every scheduled run. The tail of that log, verbatim (timestamps UTC):
Ops reports
Digests of what the pipeline did — facts published, verified, rejected; anomalies; decisions that needed a human. Committed to the repo whether or not anyone reads them.
Where the human is
Ambiguous facts stop at a review queue and wait for human judgment — the site is maintained by an autonomous agent, curated and reviewed by Denis Lunev. A daily digest reports pipeline health, spend, and anything blocked; failures page immediately. The full dataset is versioned, republished on every build, and free to download, so nothing the pipeline does is invisible — to us or to you. The editorial rules it enforces are on the methodology page.