# I went quiet for a while. This is what I was building. > Source: > Published: 2026-08-26 23:22:19+00:00 I have been quiet here for a while. This is what I was doing. theknowngood.com is live. It tracks published evaluation data for AI models: benchmark results, pricing, performance, arena ratings. Around 900 models, roughly 2,300 static pages. CSV and JSON exports are free and ungated under CC BY 4.0. No accounts, no tracking, no API. I am a systems engineer rather than a developer, so the decisions worth talking about here are all architectural. **Postgres is the source of truth, but nothing touches it at request time.** Python and Jinja2 render the whole site to flat static HTML on a schedule, so a visitor loading a model page is reading a file. That was partly about repeatability, since a static build is reproducible and diffable, and partly about cost, since flat files stay cheap under load in a way a dynamic page does not. **There are no inbound ports.** nginx binds to loopback only and the site is reached through an outbound tunnel. Nothing is listening on the public internet. That removes a category of problem instead of defending against it, which is the trade I keep making when I am the only person on call. **Every chart carries its own denominator.** Top N of M, with the real M, never the total population. That sounds like a small thing and it was the single most important correction I made to the entire project. Fair warning: this is self-hosted on hardware in my house. If it feels slow, tell me and I will go look at it. I would rather hear it from you than find it in a log three weeks later. Feedback of any kind is welcome, particularly on what I should be testing that I am not. There is an optional newsletter on the homepage, no account required. I also wrote up everything that went wrong building it, including a long stretch where most of the checks I had built were passing while measuring nothing. Link in a reply below.