Source for evolvingagentslabs.github.io.
Experiments in how agents learn, remember, and prove what they know.
index.html the experiment index
experiments/<slug>/index.html one detail page per experiment
assets/site.css all styles, hand-written
tools/{data,build}.py optional generator (see below)
Navigation mirrors labs.google/code: the index does not link to GitHub. Cards link to a detail page; the repository appears there as a CTA next to the demo, if there is one. That way a visitor reads what a thing is and what is actually established about it before landing in a source tree.
Deployment is GitHub Pages on the default branch — push to main and it is live.
There is no build step and nothing is templated; Jekyll passes the HTML through
untouched. Preview locally with:
python3 -m http.server 8000The HTML is generated but the output is committed, so the site stays buildless.
Edit tools/data.py and re-render:
python3 tools/build.py .Each entry needs a slug, name, badge, date, sort key (controls index order), the
"what if" question, a one-paragraph blurb, a repo URL, an optional demo URL, and a
list of (heading, html) sections.
You can also just hand-edit the HTML — nothing depends on the generator having
produced it. If you do, keep tools/data.py in sync or delete it.
They describe evidence, not roadmap stage. That distinction is the point of the page, and it is what makes it worth publishing next to everyone else's.
| Badge | Means |
|---|---|
Reproducible |
clone it and run it, no API key |
Results |
published findings, negative ones included |
Prototype |
runs, but needs setup or has no eval yet |
Do not promote a row without the artifact that backs it. Every detail page carries a What's proven section, and each one is expected to say plainly what is not established — the sample sizes, the missing controls, the things that break on a clean clone. That section is the reason a reader should believe the rest.
Apache 2.0 · permanently alpha