System cards

What frontier labs commit to, version by version.

Three labs tracked. Their safety frameworks (Anthropic RSP, OpenAI Preparedness Framework, Google DeepMind Frontier Safety Framework) with version-by-version diffs. Their flagship model cards with eval-category breakdowns and capability-tier classifications. Every row anchored to a primary-source URL. No interpretation. The product is structured citation.

3
Labs
6
Framework versions
0 / 6
Verified versions
3
Model cards
0 / 3
Verified cards
OpenAI
PF · PF Tier tier
Current framework
PF v2.0
effective 2025-04-15
Commitments
SeoulParis
Model cards tracked
1

Methodology.Each framework version anchors to the published PDF on the lab's canonical site. Each model card links to the lab's published system card. Anteroom does not paraphrase the frameworks or model cards. It indexes, versions, and links.

How verification works. Versions and cards marked pending are scaffolded with structurally correct metadata and source URLs, but the exact framework wording, eval-category levels, and red-team disclosures have not yet been pulled into structured form against the source PDF. Until verified, do not cite. Verification flow is in design.

Found a row that is wrong, stale, or missing? Email hello@anteroom.so.