Anthropic's Responsible Scaling Policy does not specify ASL-4 requirements until ASL-3 has been reached.

Judge as of 2024-11-11 against the text of Anthropic's Responsible Scaling Policy. True: the RSP versions in force (v1.0 of Sept 2023, and any update up to Nov 2024) say that ASL-4 commitments/requirements will be defined only later, by the time ASL-3 is reached (or before ASL-3 models are trained), and do not specify detailed ASL-4 measures. Mostly true: the RSP leaves ASL-4 largely unspecified tied to reaching ASL-3, but includes some preliminary ASL-4 content or a slightly different trigger (e.g., 'before training ASL-3 models'). Mixed: the RSP specifies ASL-4 in substantial part while leaving key elements to be defined later. Mostly false: the RSP defers ASL-4 but on a trigger unrelated to ASL-3. False: the RSP as of that date set out detailed ASL-4 requirements.

Set before the research began.

1. E1 Anthropic's Responsible Scaling Policy, Version 1.0 Anthropic · 19 September 2023 Own record

   “Capabilities and warning sign evaluations defined before training ASL-3 models”

2. E2 Anthropic's Responsible Scaling Policy, October 15, 2024 Anthropic · 15 October 2024 Own record

   “As noted, before deploying any model that passes the Capability Thresholds for the ASL-3 Required Safeguards, we will conduct a capability assessment against the forthcoming Capability Thresholds for the ASL-4 Required Safeguards.”

3. E3 Anthropic's Responsible Scaling Policy, October 15, 2024 Anthropic · 15 October 2024 Own record

   “We will update this policy with the Capability Thresholds for the ASL-4 Required Safeguards. We are currently working on defining any further Capability Thresholds that would mandate ASL-4 Required Safeguards.”

4. E4 Anthropic is Quietly Backpedalling on its Safety Commitments LessWrong (cross-post of Obsolete, by journalist Garrison Lovely) · 23 May 2025

   “While the current RSP includes high-level capability thresholds for ASL-4, it doesn't include the detailed warning sign evaluations that were part of the original commitment.”

The RSP v1.0 text says ASL-4 evaluations would be "defined before training ASL-3 models" (E1). That is a slightly different trigger from "until we've hit ASL-3". The October 2024 RSP calls the ASL-4 Capability Thresholds "forthcoming" and says Anthropic is "currently working on defining any further" thresholds that would require ASL-4 safeguards (E2, E3). So the RSP left ASL-4 mostly undefined, tied to the ASL-3 stage. However, the trigger wording differs slightly from the subject's description: v1.0 ties the ASL-4 evaluations to being defined before training ASL-3 models [E1], which fits mostly_true rather than true.