SOURCES.md lists SRC-RE-BENCH in the library: METR's measured data on AI agents doing real ML R&D tasks. The capability ladder's selfAccel bonuses (data/parameters.json -> capabilityLadder) currently cite AI-2027 and the METR task-horizon paper. If RE-Bench's findings genuinely support the bonus magnitudes, quote the figure and wire it in. The wire-a-source issue template shows the shape. This is exactly the kind of contribution the registry was built for.
SOURCES.md lists SRC-RE-BENCH in the library: METR's measured data on AI agents doing real ML R&D tasks. The capability ladder's selfAccel bonuses (data/parameters.json -> capabilityLadder) currently cite AI-2027 and the METR task-horizon paper. If RE-Bench's findings genuinely support the bonus magnitudes, quote the figure and wire it in. The wire-a-source issue template shows the shape. This is exactly the kind of contribution the registry was built for.