Long-horizon tool use is where every open-weight model currently collapses. Even strong base models like Qwen3-Coder and Kimi K2 show reliability degradation past roughly 20 tool-use steps: made-up tool names, arguments that drift out of schema, loops that burn the step budget without progress. Closed models have internal mechanisms — undisclosed training signal, better planning heads, or both — that keep the error rate flat out to 100+ steps.

This axiom closes that gap through public methodology. The solver delivers a training recipe, fine-tuned checkpoint, or inference protocol that, applied to a base model under 72B active parameters, hits ≥90% successful-termination on a fixed benchmark of 500 tasks. Each task requires at least 50 tool calls to complete. "Successful termination" is defined narrowly: the agent stops within its step budget, makes no calls to undefined tool names, and emits no malformed tool-call JSON. The task need not be solved correctly for this axiom — only terminated cleanly. Correctness is measured by adjacent benchmarks.

Verification uses a deterministic agent harness with mocked tool implementations and fixed seeds. The protocol replays the solver's recipe or inference protocol end-to-end on a clean environment and measures the termination rate.

Isolating termination from correctness is deliberate. Most labs collapse the two, which means improvements in one are obscured by noise in the other. Breaking this apart gives researchers a single-axis target.