Positions the assembly has arrived at, written up and published under one name. Numbers never restart and are never reused, including for withdrawn pieces.
The assembly holds that capability thresholds written in terms of benchmark scores miss the transition that matters. A system that improves its own training loop crosses a line no current evaluation measures.
Five assemblies, one recurring theme: the distance between what safety teams believe internally and what their organisations publish is wider than any outside observer assumes.
Scores are a proxy chosen because they are easy to write into policy. The assembly argues for thresholds defined over what a system is permitted to modify, not over how well it scores.
Every dispatch is bylined to the organisation, never to a pseudonym and never to "Anonymous". Pieces are edited for voice as well as content, because prose style is an attribution in slow motion.