Anthropic clarifies rules for updating Claude's public Constitution

Anthropic's Frontier Safety Roadmap page records a goal titled "Upholding Claude's Constitution," and the page notes that the goal's wording on updating the public Constitution was clarified on July 10, 2026. The goal had been set with a target date of October 1, 2026. On October 19, 2026, Anthropic retired it and removed it from the roadmap.

The goal covered Claude's Constitution, the public document that explains what kind of entity Anthropic wants Claude to be and what principles should guide its behavior. It had two main parts: keeping the public document in step with the version used internally, and running systematic alignment assessments of Claude's behavior. The assessments were to draw on interpretability research and be validated with deliberately misaligned test models, with findings published in system cards or Risk Reports. A separate oversight process was to review a sample of post-training data and rewards for consistency with the Constitution.

The July clarification changed the update rule. Earlier text promised an update within 90 days of relevant internal updates. The revised text says Anthropic will bring the public version in line with the latest Constitution used in training within 90 days of relevant deployments, internal or external, of models trained on updated versions. Anthropic also reserves the option of using additional guidelines and training data, such as human preference labels, without publishing them.

In its October update, Anthropic says it has kept the Constitution current and public, that alignment with it is a core metric in its automated alignment assessment, and that interpretability evidence appears in recent system cards. It adds that its oversight process surfaced nothing beyond what existing reviews caught, so it no longer runs routinely. The 90-day update commitment is to continue.