The thesis here is direct. Closed-frontier models have pulled ahead of open-weight models on agentic coding by roughly 15–25 points on SWE-bench Verified and adjacent long-horizon benchmarks. This gap is the reason most production AI deployments route through closed APIs despite the cost and lock-in. Closing it in the open would be the single most valuable event in the open-source AI ecosystem in 2026. GP-00001 is that closure, structured as a contract. An open-weight base model of at most 72 billion active parameters, under a permissive license, is fine-tuned using a fully-published training recipe. Total training compute is capped at $500,000 at list price on standard cloud. The resulting checkpoint is evaluated on a sealed slice of SWE-bench Verified and must resolve at least 70% of issues while holding inference cost at or below $0.50 per task. The verifier is a reproduction harness, not a score upload. The protocol re-runs the published training recipe on a fresh sample of the specified base checkpoint, evaluates against a held-out slice, and confirms that claimed numbers reproduce within ±2%. This is deliberately more expensive than a normal axiom verification — it is the price of the claim being worth anything. The Grand Problem decomposes into eight leaf axioms. Seven are independently valuable and payout standalone: a verified SWE-style dataset, an RL reward function for multi-file edits, tool-use reliability at 50 steps, 4× speculative decoding, bounded test-time search, large-repo retrieval, and self-verification calibration. The eighth is the integration axiom (AX-00008) that gates the $1.5M GP-gate bounty and requires the end-to-end reproduction. Formalizer attribution on the whole tree is 8%, perpetual, routed on every downstream solve. Sponsor profile is intentionally left open for a consortium structure — a foundation anchor plus cross-stakes from the three or four open-weight labs (Qwen, Kimi, DeepSeek, and peers) that stand to benefit most from the ecosystem effect. Any single sponsor can take the slot as well.

Re-run training on fresh base-model sample; evaluate held-out slice; compare within ±2%.