April 29, 2026EDITORIALlaunch · #1· 4 min read

Hello, world.

A new frontier model lands every week. The model card makes claims you can't reproduce, ships an eval table you can't audit, and goes quiet the moment the next model arrives. This blog exists to do the opposite — every post points at a real artifact in our live registry, cites specs from primary sources, and links to the verification surface so you can check the claim yourself.

tldr

This is the launch post. Going forward, each release post names a real model, states its real registry tier, lists specs we can source, and links to its live artifact page and to /verify. We don't publish numbers we can't attribute, and we don't claim a tier the registry doesn't show. Want to hear about new sealed releases? Create an account and subscribe to a publisher — one email per day, one-click unsubscribe.

post number
#1
launch date
2026-04-29
format
sourced data sheets
license
CC BY 4.0
notifications
per-publisher email
tier
EDITORIAL
audit
none (editorial)
verify
/verify

Why we're publishing this

The model-release news cycle is broken. A vendor publishes a model card with eval numbers — sometimes against a custom harness, sometimes with no harness specified at all. There's a flurry of attention, a few benchmark threads, then everyone moves on. Months later a paper shows the model was contaminated, or fails on a category nobody tested. By then it's in production, deployed on the strength of the card.

We think procurement teams, deployment engineers, and security reviewers deserve a better substrate — one with three properties:

  • Sourced. When we state a spec, it comes from the publisher's own card or paper, and we link it. When we don't have a number, we say so rather than inventing one.
  • Grounded in the registry. Every release post names a model that actually exists in our catalog, and states the tier the registry actually shows — Quarantine, Hardened, or Sealed — not a tier we wish it had.
  • Verifiable. Claims link to the live artifact page and to /verify, so the trust chain doesn't bottom out at "we said so."

What you'll see here

Mostly release posts: when a model reaches Sealed or Hardened, a short data sheet — what it is, the sourced specs, what tier it holds in the registry, and how to deploy it. Occasionally a substrate post about how tiers, witnesses, and the transparency log work. We avoid marketing for things that aren't running, vendor-quote PR, hot takes, and SEO listicles. The catalog answers "what are the best models"; the blog is for the changes to it.

What's next

The first three release posts cover models you can open in the registry right now: Qwen3-235B-A22B (Sealed), Llama-3.3-70B on Jetson Thor (Hardened), and DeepSeek-V3 (Hardened). Each links back to its live artifact so you can verify the tier and provenance independently.

Welcome.

If you have to defend a model selection to an ATO board, need to know what runs on an edge box, or just care about open-weight AI being deployed responsibly — this is for you. See you on the next post.

more releases