The $0.50 per task inference budget in GP-00001 is tight. At current open-weight inference costs on an 8×H100 node, a 72B reasoning model running SWE-bench at full-length traces lands somewhere between $0.80 and $2.00 per task depending on trajectory length. Closing that gap without compromising model quality is the dominant inference-side research question for open labs right now.

This axiom delivers that closure. The solver publishes a speculative decoding scheme — draft model, verification protocol, acceptance rule, and any auxiliary components — along with code and weights. The scheme must achieve at least 4× decode throughput on a reference 72B reasoning model, measured under a standardized benchmark harness on an 8×H100 node. Quality degradation, measured as mean performance across ten standard reasoning benchmarks (MATH, GSM8K, AIME, LiveCodeBench, SWE-bench Lite, HumanEval+, BBH, DROP, MMLU-Pro, FrontierMath-Lite), must not exceed 0.5% absolute.

Verification runs the published scheme end-to-end. The protocol reproduces the throughput measurement on a standardized rig and replays the benchmark suite on both the baseline and accelerated configurations. Axiom passes when the throughput ratio is ≥4.0 and the quality delta across all ten benchmarks averages ≤0.5%.

The downstream value is not only AX-00008. Every open lab running 72B-class reasoning models in production has this problem, and a solution that works on one base architecture typically transfers.