World-3 · web and tooling · World-3 collaboration. gpt-6-astra on Pi, 26 agents, connected, discovery prompt: 1,319 sessions, 97 agent hours, 0.44 B tokens, $1,024 at list prices. From data/per-release-v3/41.5 in the paper’s figures repository at 9014a8cc, fetched from self-evolving/experiment-41-5 at d992015d on Hugging Face. The states and releases come from the checkpoint self-evolving/experiment-41-5 @ c446b3f0 (branch main). The score is rebuilt from the forge’s history, one state every two hours, each task carried at its last measured value.
Scarry
1.017
+1.7% over the seed at the last reading, after 1,286 sessions
+0.1 pp from the previous reading
Sstrict
1.017
a package with a failed check or benchmark scores 1.0
Movers
+12 / −0
tasks beyond their noise floor at the last reading
Readings
5
latest Sep 24, 13:39 UTC
Run
1,319 sessions
completed work sessions
eval set v3.x · speedup over the seed
Scarry at each reading
64 releases measured
Breakdownafter 1,286 sessions · Sep 24, 13:39 UTC
Scarry 1.017 · Sstrict 1.017 · +1.7% over the tasks with a value · 20/20 tasks measured