Runs, instance by instance.

Every run in the paper

Each row is one student model on one track. Every learner in it plays the stream and then the two frozen tests, and is read against itself reset. GPT-6 Luna's runs replay here. The others' records and trajectories go public with the paper, in the same repository.

Start here

• Prime Agent, medium effortThe digest that opens every session, which said “No saved harness entries yet.” every time, returning tasks included.
• Prime Agent, high effortIts reasoning once the task was solved, all of it about finishing: “It looks like the user wants me to finish the action since they've indicated that it's solved.”
• Hermes Agent, low effortThe 11 skills it held at the end of its stream, none of them named for a task.
• OpenClaw, asked to storeNotes written under names it made up afresh each time, which before acting it almost never looked for when a task recurred.
• Prime Agent, told to keep a program libraryA returning task's kept program, run first on the new input, as the program library does.

Pick a run

All 114 of GPT-6 Luna’s runs

Budget used is the share of the cap of 16 that an instance was charged, averaged over the instances, where a solved instance costs 1 for each wrong answer and each demonstration and an unsolved one all 16, so lower is better. Solved is the share solved within the cap. The stream is its 200 instances. After it, a run on the stream went on to the epilogue, the 100 held-out tasks played while the learner could still learn, which the frozen tests replace: the floors, the program library and Prime Agent played all of it, and OpenClaw and Hermes Agent were stopped part of the way in, some at its start. A fresh run serves those 100 tasks alone, with the learner reset before each.