Revisions · Safety Card Ledger

Summary

The ledger records 15 revisions to 9 documents by Anthropic, Google DeepMind, OpenAI and xAI, dated between Dec 2025 and 9 Sep 2026. Of these, the document's own changelog explains 7, partly explains 3 and does not explain 2; 3 are in documents that keep no changelog. In 3 of them, a document was marked updated and there is no record of what changed.

Recorded revisions by explanation status
Explanation	Revisions	Developers
ExplainedThe document's changelog describes the change.	7 — show the revisions marked Explained	Anthropic, OpenAI
Partly explainedThe changelog mentions the change, but not in full or without a reason.	3 — show the revisions marked Partly explained	xAI
Not explainedChanged without an explanation in the document's changelog.	2 — show the revisions marked Not explained	xAI
No changelogThe document keeps no changelog.	3 — show the revisions marked No changelog	Google DeepMind
All statuses	15	Anthropic, Google DeepMind, OpenAI, xAI

The ledger holds the revisions recorded so far. Developers' changelogs also date 25 later versions of documents whose changes are not recorded in the ledger; each document's page lists its versions.

Every revision, newest first

Grouped by the revision of a document they belong to, and dated by the later version as the document states it, not by when we read it. Select a value for its source and its checks.

KeyVerifiedUnverifiedDisputedCorrected read off a figure or stated in wordsA value opens its source and history.

Developer

Explanation

Metric family

Showing all 15 revisions.

9 Sep 2026
GPT-6 Astra System Card
OpenAI · System card · 1 change
Earlier version
3 Sep 2026, dated by our source registry. Known to exist; no copy held
Later version
9 Sep 2026, dated by the document's changelog. Known from the changelog; no copy held
1. Wording changed
  Alignment section clarified and its limitations expanded; section on verbalized metagaming and oversight gaming renamed and revised
  Explained — changelog describes the changes
  How we know Re-read from the change log of the copy retrieved on 27 Sep 2026: two entries dated 9 Sep. The change log gives only the section's current title. An extraction note on m-001261 says the section was titled 'Metagaming and Alignment Faking' before; we cannot confirm that without a copy of the version of 3 Sep.
  How it was checked Confirmed from changelog
19 Aug 2026
GPT-5.6 Preview System Card
OpenAI · System card · 1 change
Earlier version
25 Jun 2026, dated by our source registry. Known to exist; no copy held
Later version
19 Aug 2026, dated by the document's changelog. Known from the changelog; no copy held
1. Value changed
  GPT-5.5 hard-negative protein-binding prediction, pass@4
  ~~0.4%~~1.5%
  Explained — earlier figure was the pass@1 score
  How we know Changelog read 26 Sep 2026 (no v0 value row in this document)
  How it was checked Confirmed from changelog
19 Aug 2026
GPT-5.6 System Card
OpenAI · System card · 1 change
Earlier version
9 Jul 2026, dated by our source registry. Known to exist; no copy held
Later version
19 Aug 2026, dated by the document's changelog. Known from the changelog; no copy held
1. Value changedM10 · Dangerous capabilities and risk determinations
  GPT-5.5 hard-negative protein-binding prediction, pass@4
  ~~0.4%~~1.5%
  Explained — earlier figure was the pass@1 score
  New value
  How we know Changelog read 26 Sep 2026; same entry also appears in the GPT-5.6 Preview card changelog
  How it was checked Confirmed from changelog
17 Aug 2026
Grok 4.6 Model Card
xAI · Model card · 5 changes
Earlier version
12 Aug 2026, dated by our source registry. Version of 12 Aug 2026 (opens an external site) · an earlier copy, hosted elsewhere
Later version
17 Aug 2026, dated by the document's changelog. Version of 17 Aug 2026 (opens an external site) · retrieved 26 Sep 2026
1. Value changedM7 · Harmful compliance and over-refusal
  HackerBench v0.2 harmful/dual-use compliance, Grok 4.6 (high)
  ~~16.7%~~6.9%
  Partly explained — changelog lists "corrected eval results" for this eval; no reason
  Old value New value
  How we know The old value was read in the earlier copy and the new one in the later copy. Blind check 26 Sep 2026 (both versions)
  How it was checked Blind-verified
2. Value changedM5 · Honesty and hallucination
  MASK dishonesty, Grok 4.6 (high)
  ~~3.8%~~1.90%
  Partly explained — listed as corrected; no reason
  Old value New value
  How we know The old value was read in the earlier copy and the new one in the later copy. Blind check 26 Sep 2026 (both versions)
  How it was checked Blind-verified
3. Value changedM7 · Harmful compliance and over-refusal
  Self-harm compliance, Grok 4.6 (high)
  ~~3.7%~~0.84%
  Partly explained — listed as corrected; no reason
  Old value New value
  How we know The old value was read in the earlier copy and the new one in the later copy. Blind check 26 Sep 2026 (both versions)
  How it was checked Blind-verified
4. AddedM7 · Harmful compliance and over-refusal
  FORTRESS-RN R/N refusal accuracy added, Grok 4.6 (high)
  Not in the earlier version97.9%
  Not explained — Added without a changelog note.
  New value
  How we know The value was read in the later copy; the earlier copy does not have it.
  How it was checked Not yet checked
5. Wording changedM10 · Dangerous capabilities and risk determinations
  Bio capability lift statement relative to Grok 4.5, Grok 4.6
  ~~no appreciable lift~~noted in the biological domain but is limited
  Not explained — The changelog does not mention this change.
  Old value New value
  How we know The old value was read in the earlier copy and the new one in the later copy. The FORTRESS-RN addition in the same revision is recorded separately.
  How it was checked Not yet checked
3 Aug 2026
GPT-5.6 System Card
OpenAI · System card · 2 changes
Earlier version
9 Jul 2026, dated by our source registry. Known to exist; no copy held
Later version
3 Aug 2026, dated by the document's changelog. Known from the changelog; no copy held
1. AddedM9 · Prompt injection
  GPT-Red indirect prompt injection attack success rate added, GPT-5.6 Sol
  Not in the earlier version3.77%
  Explained — The changelog says the GPT-Red prompt-injection results were added.
  New value
  How we know From the document's changelog, as summarised in our source registry.
  How it was checked Not yet checked
2. AddedM9 · Prompt injection
  GPT-Red instruction-hierarchy (direct prompt injection) attack success rate added, GPT-5.6 Sol
  Not in the earlier version0.051%
  Explained — The changelog says the GPT-Red prompt-injection results were added.
  New value
  How we know From the document's changelog, as summarised in our source registry.
  How it was checked Not yet checked
25 Jun 2026
Claude Fable 5 & Claude Mythos 5 System Card
Anthropic · System card · 1 change
Earlier version
11 Jun 2026, dated by the document's changelog. Known from the changelog; no copy held
Later version
25 Jun 2026, dated by the document's changelog. Known from the changelog; no copy held
1. Wording changed
  Alignment risk in executive summary
  ~~low~~very low
  Explained — corrected to match §2.4
  How we know Changelog read 26 Sep 2026
  How it was checked Confirmed from changelog
17 Jun 2026
Claude Opus 4.8 System Card
Anthropic · System card · 1 change
Earlier version
3 Jun 2026, dated by the document's changelog. Known from the changelog; no copy held
Later version
17 Jun 2026, dated by the document's changelog. Known from the changelog; no copy held
1. Value changedM10 · Dangerous capabilities and risk determinations
  Long-form virology task 2 end-to-end score, Claude Opus 4.8
  ~~0.89~~0.90
  Explained — The changelog lists the corrected score.
  New value
  How we know Old and new text as the source registry quotes the changelog entry of 17 Jun 2026; v0 row 350 records the corrected value. The changelog has not been re-read by us, so the revision is unverified.
  How it was checked Not yet checked
May 2026
Gemini 3 Pro Model Card
Google DeepMind · Model card · 1 change
Earlier version
18 Nov 2025, dated by our source registry. Known to exist; no copy held
Later version
May 2026, dated by the document header. Version of May 2026 (opens an external site) · retrieved 26 Sep 2026
1. Marked updated
  Card marked "Last updated May 2026"
  The document was marked updated. There is no record of what changed.
  No changelog
  How we know Header read by extraction pass
  How it was checked Not yet checked
May 2026
Gemini 3.1 Flash-Lite Model Card
Google DeepMind · Model card · 1 change
Earlier version
3 Mar 2026, dated by our source registry. Known to exist; no copy held
Later version
May 2026, dated by the document header. Version of May 2026 (opens an external site) · retrieved 26 Sep 2026
1. Marked updated
  Card marked updated May 2026
  The document was marked updated. There is no record of what changed.
  No changelog
  How we know Header read by extraction pass
  How it was checked Not yet checked
Dec 2025
Gemini 2.5 Flash Model Card
Google DeepMind · Model card · 1 change
Earlier version
26 Sep 2025, dated by our source registry. Known to exist; no copy held
Later version
Dec 2025, dated by the document header. Version of Dec 2025 (opens an external site) · retrieved 26 Sep 2026
1. Marked updated
  Card marked updated Dec 2025
  The document was marked updated. There is no record of what changed.
  No changelog
  How we know Header read by extraction pass
  How it was checked Not yet checked

Source: Safety Card Ledger v0.1 · Data from developer system cards · CC BY 4.0