An archived edition of vanemmerik.ai/Signal, preserved exactly as published — including that day's generative art seed.
That day's lede
✓ Fact-checked
On Sept. 18, Anthropic named Accenture's applied-AI unit Faculty as its first "embedded evaluator" — a third-party safety assessor stationed inside Anthropic with access comparable to an employee's, tasked with red-teaming models, running alignment assessments, testing safeguards, and reporting findings independently rather than through an outside audit. Anthropic and Accenture each expect to invest at least $1 billion over the next five years to build out the program, and Anthropic said it's in talks with METR and other nonprofits about piloting elements of embedded evaluation with their own funding, with more evaluator partnerships expected in the coming weeks. It remains the most substantial Anthropic news of the past several days — the freshest Claude Code release (Sept. 19) was a narrower billing-mechanics change — so this partnership, alongside today's brief on Claude Code's redesigned Projects and Sept. 17's life-sciences and AI-pace announcements, are this morning's most worthwhile reads.
Claude Code · Sep 19
Version 2.1.278, released Sept. 19, changes auto mode's default behavior on the Claude API, Enterprise, Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry, and gateway setups: instead of Claude Code running its own billed classifier requests to screen risky actions before they execute, the server now performs those checks as part of the session's existing model requests at no extra charge. Pro, Max, and Team users were unaffected since auto mode was already free for them; sessions behind an LLM gateway that strips or rewrites request data may still fall back to the old, billed client-side classifier and see a one-time notice, with CLAUDE_CODE_AUTO_MODE_SERVER=0 available to opt out entirely. The new /status now includes an "Auto mode server" row showing which mode a session is using.
Claude Code · Sep 17
Announced Sept. 17, Claude Code Projects is being rebuilt around a single conversation instead of a shared folder: you describe what needs doing, and Claude scopes the request, delegates work across parallel threads — each its own Claude Code cloud session working on its own branch and copy of the repo — then reviews outputs and assembles the result. A shared-memory layer carries decisions, release schedules, and team preferences across sessions so context doesn't need re-explaining, and a project library collects files and artifacts for later reuse. It's in beta today for select Pro and Max users on Claude Code cloud sessions, with local-machine execution and wider Team/Enterprise availability coming soon.
Anthropic · Sep 17
On Sept. 17, Anthropic launched the Life Sciences Verification Program in beta, letting vetted academic labs, biotechs, and pharmaceutical companies apply for access to Claude Mythos, Opus, and Sonnet models with biology safeguards more permissive than the generally available Fable models — unlocking work like drug discovery, clinical development, and manufacturing that's currently blocked. Applicants are reviewed for research credentials, security standards, and ethical oversight, and get a "standard use" grant for most life-sciences work or a "high-risk use" add-on that removes life-sciences-specific safeguards for dual-use projects; Anthropic says the program shifts from real-time blocking toward offline monitoring of usage against declared use cases. It's live now on the API and for Enterprise and Team plans, with individual-plan access coming later.
Anthropic Institute · Sep 17
Also on Sept. 17, Anthropic's policy institute proposed three measurements meant to give outsiders visibility into frontier AI progress that capability benchmarks don't capture: an "AI R&D automation index" rating how much of Anthropic's own AI research is done by AI (Claude now leads an estimated 26% of that work, Anthropic says, up from under 1% in February); agent-oversight metrics tracking roughly 30,000 concurrently running agents, of which Anthropic says 100% of actions pass through an online monitor before executing, with about 0.002% of over a billion August decisions blocked by that monitor; and a compute-allocation metric — a July snapshot found about 6% of Anthropic's AI R&D compute went to safety work. Anthropic is proposing labs report figures like these publicly and consistently, to make claims about AI's pace of development checkable rather than anecdotal.