selekoa

A documentary history of machine intelligence.

Found and cross-checked by machines · How this works →

Comparison · 18 questions · 7 frameworks

Compare frameworks

What does each developer’s safety framework say on the same questions, edition by edition?

Each row is one question, worded so that it fits every framework. Each column is one framework, showing its latest edition unless you choose another. A cell gives what that edition says on the question, with the page, or records that no passage was found in a stated search. Six questions give the edition’s own words; the other twelve give Selekoa’s paraphrase, marked as such. The questions are Selekoa’s. An edition’s own terms, such as “ASL” or “Critical Capability Level”, appear only in quotation marks.

A status records what an edition states. It does not show implementation, compliance or effectiveness, and a shared row does not make two provisions equivalent.

Download every cell as CSV

What the statuses mean

Stated
A passage meets the question’s criterion as worded, however conditional or narrow its scope. The condition or scope is in the note.
Stated, partly
A passage has what the criterion asks for, but in a weaker form than the criterion’s own words: for example, no named person or body where the question asks for one, or a “may” where it asks for a statement that something will happen.
Not found in the recorded search
No passage meeting the criterion was found. The record lists the sections read, the search terms and every passage set aside, with the reason. It does not show that the developer lacks the practice, or that a provision was deleted.
Does not apply
The question does not apply: the edition is the first in its framework, so there is no earlier edition to change from.
Limits of this comparison
  • Conditions that matter are in each cell’s note. A “stated” cell can carry conditions as important as those of a “stated, partly” one.
  • The page compares and does not rank. There are no scores or totals, and columns are in a fixed order.
  • Six questions, on required actions, who decides and exceptions, quote the edition. The other twelve, on risk domains, thresholds, evaluation timing, reporting and revision, give a paraphrase with the page, because quoting them as well would exceed Selekoa’s quotation limit for most editions. A paraphrase states what the passage says, in other words; read the edition at that page before relying on it.
  • Passages were selected, checked and placed under each question by machine, and checked by a reviewer from a different model family. The maintainer decided the borderline cases.
Editions shown: the latest of each framework
Edition shown for each framework

Risk domains coveredOther named domains

Criterion: Any further domain the edition names as tracked (for example misalignment, deceptive behaviour, sabotage of oversight). Research categories or monitored-only domains are recorded with their status as worded. Overlap rule: loss of control or misalignment defined through deception or undermining human oversight is recorded here; autonomy, self-replication and AI research and development go under “Autonomy, self-improvement or AI research and development”.

  1. Anthropic · Responsible Scaling Policyv3.4 · effective 2026-07-08

    Edition textArchived copy, captured 2026-07-10

    Stated

    • Paraphrase: The edition’s threshold table has a row titled “Misaligned AI systems in high-stakes settings”.

      PDF p. 7, §1, threshold table, row title, left column (“Capability or usage threshold”)

    Note: A threshold row on sabotage by AI systems. Its industry-wide column calls for compartmentalization, monitoring, capability assessments and evidence on whether models tend to deceive, manipulate or sabotage their users.

    Inspect sources and checks

  2. Google DeepMind · Frontier Safety Frameworkv3.1 · published 2026-04-17

    Edition textArchived copy, captured 2026-04-23

    Stated

    • Paraphrase: The edition says Google DeepMind has determined, on early research, that significant or severe risks from future models may be likeliest in CBRN, cyber, harmful manipulation, and in “machine learning R&D and misalignment”.

      PDF p. 5 (printed p. 5), §1.3.1 Risk Identification, fourth sentence

    Note: Misalignment is no longer a separate exploratory section; it is part of the combined “ML R&D and Misalignment” domain, tracked through the “Stealth and Situational Awareness TCL”, which the edition describes (PDF p. 14) as a baseline capability level above which misalignment risks may start to arise.

    Inspect sources and checks

  3. Meta · Frontier AI / Advanced AI Scaling Frameworkv2.0 · published 2026-04-08

    Edition textArchived copy, captured 2026-04-08

    Stated

    • Paraphrase: The edition names the outcome “Loss of Control 2” as losing the ability to monitor how AI behaves in operation, with the threat scenario that AI disables or corrupts monitors so as to reduce oversight and achieve its objectives.

      PDF p. 23, §3.4.3 Loss of Control, table row 'Loss of Control 2', Outcome and Potential Threat Scenario columns

    • Paraphrase: The edition says physical autonomy is now a further, nascent outcome of Frontier AI.

      PDF p. 36 (printed p. 36), §4.2.4 Emerging Evaluations and Outcomes, 'Physical autonomy', first sentence

    Note: Meta names one domain, Loss of Control, one of the three areas the edition says the Framework currently focuses on (PDF p. 1). Under the “Risk domains covered” overlap rule its outcome Loss of Control 2, defined through AI undermining monitoring, is recorded here; Loss of Control 1 (autonomous AI research and development) is recorded under “Autonomy, self-improvement or AI research and development”. Physical autonomy (second passage) is listed under 'Emerging Evaluations and Outcomes' with a sample outcome and threat scenario, not incorporated into the thresholds.

    Inspect sources and checks

  4. Microsoft · Frontier Governance FrameworkFeb 2026 · revised 2026-02

    Edition textArchived copy, captured 2026-04-24

    Stated

    • Paraphrase: The edition lists “Loss of control” as a tracked high-risk capability: the ability of a model to undermine effective control by humans through adaptive, deceptive or self-reinforcing means, so that once deployed it can no longer reliably be directed, modified or shut down.

      PDF p. 4 (printed p. 3), §2 Capabilities and risks, 'Tracked high-risk capabilities', fourth bullet

    Note: A separately named tracked capability, recorded here under the “Risk domains covered” overlap rule because the edition defines it through deceptive mechanisms that undermine human control. No thresholds are set: Appendix I says Microsoft is researching evaluation approaches and risk acceptance criteria for loss of control risk (PDF p. 15).

    Inspect sources and checks

  5. OpenAI · Frontier Governance FrameworkMay 2026 edition · published 2026-05-28

    Edition textArchived copy, captured 2026-05-29

    Stated, partly

    • Paraphrase: The edition says loss-of-control risks may arise from misalignment with what humans intend or value, from self-improvement, from deception by the model, or from autonomous self-improvement.

      PDF p. 12 (printed p. 10), §2.4 Risk tiers › 'Loss of control', first paragraph, second sentence

    Note: No fifth category is named; misalignment and deception are named as sources of loss-of-control risk, not as separate categories.

    Inspect sources and checks

  6. OpenAI · Preparedness FrameworkVersion 2 · published 2025-04-15

    Edition textArchived copy, captured 2025-04-15

    Stated

    • Paraphrase: The edition lists Undermining Safeguards as a Research Category: a model's ability and propensity to act so as to undermine the safeguards placed on it.

      PDF p. 8 (printed p. 7), Table 2 'Research Categories', fourth row

    Note: Named as a Research Category, not a Tracked Category: the edition sets no thresholds for it. Table 2 also names Sandbagging as a Research Category; deception appears as an example within Undermining Safeguards.

    Inspect sources and checks

  7. xAI · Frontier AI FrameworkJune 2026 edition · effective 2026-06-30

    Edition textArchived copy, captured 2026-07-16

    Stated, partly

    • Paraphrase: The edition says xAI focuses in particular on malicious-use risk and on the risk of safeguards being circumvented, giving jailbreaks as an example.

      PDF p. 1, §1 Introduction, third paragraph, first sentence

    • Paraphrase: The edition says that, within its risk management framework, xAI may also identify and assess further risks.

      PDF p. 1, §1 Introduction, third paragraph, final sentence

    Note: No domain beyond the four is named as tracked; safeguard circumvention is named as a focus alongside malicious use, and other risks only as a possibility.

    Inspect sources and checks