Until today the verification record had a gap we had been honest about but had not closed: ten entries were still carrying August dates. This pass checked all 23, and asked of each the question the new method rule demands — not just is this field still right but is this still the thing we say it is.
Four findings. Two products had moved without announcing it: We were listing an agent that had been dead for five months and Open Interpreter's repository is no longer Open Interpreter. One licence was wrong as a consequence of the second. And one surface was wrong on its own merits — UI-TARS was recorded as controlling both a desktop and a browser, but its repository documents desktop and mobile use and points at separate sibling projects for the desktop application and for web automation. Browser control is not that repository's job, so the column now reads Desktop.
Three smaller things that did not change a field but belong on the page. Cua is not uniformly MIT — Cua Spaces is source-available under FSL-1.1-MIT — and the project says plainly that one of its own packages has known security issues and should not be run. Agent S's README now leads with a hosted commercial product as the production version of the same technology. Skyvern's AGPL covers the core but not the anti-bot code in its managed cloud.
And one shift in tone rather than status: OpenAI's guide for Computer use tool now recommends code execution for its newest model and describes the computer tool as remaining supported as an alternative. Still GA, still documented — but no longer the route being pushed, which is usually how these things begin.
Coverage, stated plainly, because a verification claim without coverage is marketing: 15 entries primary-sourced, 6 secondary — the OpenAI, Meta and Perplexity entries, whose vendors refuse automated access — and 2 deliberately unconfirmed, where we believe a product is finished but no vendor has said so.