Open data

Threshold Watch

Every frontier lab publishes a safety framework: the capability thresholds that are supposed to trigger stronger safeguards. None of them publishes a version history. The thresholds differ between labs, change without notice, and changes are how commitments soften. This page does the versioning for them: each framework is fetched on a cadence, normalized, hashed, and archived. When the content changes, the new snapshot lands here, dated, with a diff where the source is text.

This is commitment one of the Control Stack Compact run in reverse: thresholds should be published in a common format and versioned like a changelog by the labs themselves. Until they are, the changelog lives here.

Subscribe to changes: RSS feed (a new item means a framework changed or entered tracking). Related: the Frontier Incident Timeline and the Compact Clauses for procurement.

Commitment one, scored
LabFramework publishedFetchable by toolingVersioned by the labDiffable changelog
AnthropicYesYesYes (v3.4)Announcements only
OpenAIYesPDF only; page blocks toolingYes (v2)No
Google DeepMindYesYesYes (3.1)Blog summaries only
xAIYesYesDated editionsNo
MetaYesNo; PDF rejects automated retrievalYes (v2.0)No
AmazonYesYesUnclearNo

Six labs publish a framework. None publishes a changelog a third party can verify. That last column is what this page exists to fill.

Anthropic

Responsible Scaling Policy

current: v3.4, effective July 8, 2026
Tracked source: https://www.anthropic.com/rsp (page text, tracked with diffs)
Tracked as rendered page text.
snapshotsha256bytesfiles
2026-09-08 (first)860f164836fea18c24,960raw
Last checked 2026-09-08 · status: ok · no changes observed yet
OpenAI

Preparedness Framework

current: v2, last updated April 15, 2025
openai.com blocks automated retrieval of the framework's HTML page; the canonical PDF on OpenAI's CDN is tracked byte-for-byte instead.
snapshotsha256bytesfiles
2026-09-08 (first)fae6e4cc6f7212a0170,399raw
Last checked 2026-09-08 · status: ok · no changes observed yet
Google DeepMind

Frontier Safety Framework

current: 3.1, as of April 17, 2026
Tracked byte-for-byte from DeepMind's published PDF.
snapshotsha256bytesfiles
2026-09-08 (first)ad10d2d240e0d557447,155raw
Last checked 2026-09-08 · status: ok · no changes observed yet
xAI

Frontier AI Framework

current: June 2026 edition, effective June 30, 2026
xAI publishes each version at a new URL, so a framework change shows up here as this document going stale rather than its hash changing.
snapshotsha256bytesfiles
2026-09-08 (first)2c3c6313bd2fb6ee203,142raw
Last checked 2026-09-08 · status: ok · no changes observed yet
Meta

Advanced AI Scaling Framework

current: v2.0, dated April 7, 2026
Meta's framework PDF rejects automated retrieval. The fetch failure below is left visible on purpose: a threshold document tooling cannot read is part of the finding.
snapshotsha256bytesfiles
Last checked 2026-09-08 · status: error: HTTPError
Amazon

Frontier Model Safety Framework

Tracked as rendered page text.
snapshotsha256bytesfiles
2026-09-08 (first)513f9ba04df1f86530,392raw
Last checked 2026-09-08 · status: ok · no changes observed yet
Method

Read-only fetches of each lab's canonical public document. HTML sources are reduced to visible page text before hashing so cosmetic markup changes don't register as framework changes; PDFs are hashed byte-for-byte. A snapshot is kept only when the hash changes. Fetch failures are shown as-is rather than papered over. Source code ships in the major-labs repository (threshold-watch/fetch.py).