Recovering all 119 Feynman equations on a laptop: a full score on the strict SRBench ruler at about two seconds per problem

SparkOrigin recovers all 119 equations in the SRBench Feynman benchmark, scored by SRBench's own strict symbolic-solution criterion (La Cava et al., 2021) on the official public data and sampling ranges, at roughly two seconds per problem on a laptop CPU, in pure NumPy, with no neural network. Under the same strict ruler the strongest published methods reach 53–56% at CPU-hour budgets, so the result is a point on the accuracy–efficiency frontier rather than accuracy alone. We report three numbers rather than one: 119/119 numerically (R² > 0.999 inside and beyond the sampled range); 119/119 symbolically under the SRBench rule with two disclosed normalizations (physical positivity and exact π/√ constant recognition); and 81/119 under a naive plain-symbol reading — the 38-problem gap being provably equivalent forms, each one listed. On held-out novel forms (five independently authored blind sets, 260+ unseen problems) recovery is lower (~78% on typical problems, ~64% on hard ones), with zero false positives. This deposit contains the paper (English and Chinese) and a reproduction package that re-verifies all 119 equations WITHOUT the discovery engine, using a line-for-line port of SRBench's ruler and the public data. The discovery engine is proprietary and is not distributed; its live operation is shown in an accompanying timestamped screencast. This is the clean, noise-free benchmark; robustness on noisy and real-world data is a separate axis and is not claimed here.