Anthropic's reset provable-inference deadline lapses without a public update

Anthropic first gave itself 43 days for Phase 1, from April 2 to May 15 ^[1]. It moved the target ten days before that date came due ^[3]. The new window ran 181 days ^[4], a little over four times the first ^[5]. Forkast logs one more change to the record, a typo correction on July 29 ^[15].

For a milestone like this, the first question is what evidence of completion would look like. The deliverable is an inventory of components, costs and timelines ^[6]. Finished, it is a planning document. Unless Anthropic publishes it or says it exists, an outsider can test only the calendar. Forkast's test is the absence of a blog post, press release or social media update by close of business on Sept 30 ^[4].

That test is weaker than it looks. Anthropic's Responsible Scaling Policy page says the planned moonshot projects have launched and that the initial launch goals were replaced with more detailed objectives for ongoing work ^[9]. Forkast itself allows that silence does not equate to failure, and that it may reflect a change in internal reporting or a recalibrated scope ^[10]. The thing this doesn't tell you is whether the replacement objectives still hold a Phase 1 date, or what that date now is.

I think the lapse is weak evidence about the engineering and fair evidence about the policy's format. When the deliverable stays inside the company and the goals can be rewritten on the policy page, an outside reader cannot tell a missed deadline from a retired one.

The engineering constraint is older than the deadline. The June 2025 Confidential Inference authors noted that not all hardware accelerators fully support confidential computing environments ^[8]. Their approach tied trust to hardware-rooted attestation through Trusted Platform Modules ^[7], so signing outputs to specific weights needs chips that can produce that attestation. A Phase 1 plan of components and costs is where that gap would have to be costed.

Pawan Khandavilli, in an analysis Forkast cites, argued that the harder problem lies past inference: the industry, he wrote, lacks a framework for confidential agency, and agent identity today rests on trusting API keys and system prompts ^[12]. He listed four gaps: hardware-rooted agent identity, measured policy-as-code, portable attestation evidence and cross-cloud federation ^[13].