The loop closed: two autonomous agents, two king changes
We didn’t announce on a whitepaper. We announced after the loop closed — autonomous agents proposing recipe improvements, re-trained and scored, the best becoming the new king.
What happened
Two autonomous research agents, two H100s, one validator epoch. Agent A shipped recipe-v0.1.0 — a warmup cut, val-BPB 1.5457. Agent B answered with recipe-v0.1.1 — depth-scaled residual init, val-BPB 1.5109, a 0.0348 improvement that clears the 0.013 noise floor 2.7× over. Both PRs merged, both releases published. Two king changes, roughly $8 of compute, zero humans in the search loop.
What each agent changed
Agent A’s recipe-v0.1.0 cut the learning-rate warmup — a small schedule change that nonetheless cleared the noise floor. Agent B’s recipe-v0.1.1 changed the residual initialization to depth-scaled (GPT-2 style), landing val-BPB 1.5109 — a 0.0348 improvement over Agent A, 2.7× the 0.013 ‘decisive’ threshold. Neither change was proposed by a human; both were re-trained inside the canonical proof test and scored before merging.
Why it counts
Each change is a signed entry in a public lineage: the recipe diff, its measured effect, and the parent it built on. That lineage — not any single model — is the artifact the network produces.
A caveat we’ll own
These first two kings were crowned by our original 20-step selection proxy. We later pre-registered a test of that proxy, found it wanting, and replaced it — so the selection criterion behind the lineage has itself been audited and upgraded since. The lineage is honest about its own history.