Andrew Fribush
fig. 2.2.7 · the selector's query time against the fastest explicit route, all 36 workloads.
PER QUERY, THE SELECTOR WAS NEVER SLOWER THAN 1.025× THE BEST ROUTE YOU COULD FORCEPER QUERYFIRST USE, SETUP INCLUDEDONE DOT IS ONE TEST CELL: THE SELECTOR'S MEDIAN TIME OVER THE BEST EXPLICIT ROUTE'S,LOG SCALE. LIT: IT CHOSE SCREENING (18 OF 36).EXISTING CELLS 188 SCREENEDARXIV0.800.900.95COCO0.800.900.95GLOVE0.800.900.95ISOTROPIC0.800.900.95CLUSTERED0.800.900.95IDENTICAL0.800.900.950.33×0.5×0.75×1×1.25×1.5×← FASTERSLOWER →arXiv 0.80: 0.62× per query, 3.0× first use; chose screening (rank128)arXiv 0.90: 1.002× per query, 2.6× first use; chose screening (rank64)arXiv 0.95: 1.00× per query, 2.4× first use; chose screening (rank64)COCO 0.80: 1.00× per query, 2.6× first use; chose denseCOCO 0.90: 1.005× per query, 2.8× first use; chose screening (rank64)COCO 0.95: 1.004× per query, 2.7× first use; chose screening (rank64)GloVe 0.80: 1.006× per query, 2.1× first use; chose denseGloVe 0.90: 1.00× per query, 2.1× first use; chose denseGloVe 0.95: 1.000× per query, 2.1× first use; chose denseisotropic 0.80: 1.00× per query, 2.7× first use; chose denseisotropic 0.90: 0.44× per query, 3.1× first use; chose screening (rank128)isotropic 0.95: 0.69× per query, 2.8× first use; chose screening (rank128)clustered 0.80: 1.023× per query, 2.7× first use; chose denseclustered 0.90: 1.022× per query, 2.7× first use; chose denseclustered 0.95: 0.39× per query, 2.9× first use; chose screening (rank128)identical 0.80: 1.023× per query, 35× first use; chose denseidentical 0.90: 1.00× per query, 35× first use; chose denseidentical 0.95: 1.00× per query, 35× first use; chose denseHELD-OUT CELLS 1810 SCREENEDARXIV0.850.9250.975COCO0.850.9250.975GLOVE0.850.9250.975ISOTROPIC0.850.9250.975CLUSTERED0.850.9250.975IDENTICAL0.850.9250.9750.33×0.5×0.75×1×1.25×1.5×← FASTERSLOWER →arXiv 0.85: 0.54× per query, 2.8× first use; chose screening (rank128)arXiv 0.925: 1.002× per query, 2.5× first use; chose screening (rank64)arXiv 0.975: 1.003× per query, 2.4× first use; chose screening (rank64)COCO 0.85: 0.94× per query, 3.1× first use; chose screening (rank128)COCO 0.925: 1.008× per query, 2.7× first use; chose screening (rank64)COCO 0.975: 1.003× per query, 2.7× first use; chose screening (rank64)GloVe 0.85: 1.000× per query, 2.1× first use; chose denseGloVe 0.925: 1.00× per query, 2.0× first use; chose denseGloVe 0.975: 1.00× per query, 2.1× first use; chose denseisotropic 0.85: 1.018× per query, 2.7× first use; chose denseisotropic 0.925: 0.34× per query, 3.1× first use; chose screening (rank128)isotropic 0.975: 1.020× per query, 2.8× first use; chose screening (rank128)clustered 0.85: 1.005× per query, 2.8× first use; chose denseclustered 0.925: 0.46× per query, 3.0× first use; chose screening (rank128)clustered 0.975: 0.79× per query, 2.8× first use; chose screening (rank128)identical 0.85: 0.97× per query, 36× first use; chose denseidentical 0.925: 1.025× per query, 35× first use; chose denseidentical 0.975: 1.00× per query, 36× first use; chose denseWORST 1.025×0.34×, RANK-128 SCREENING

source: gpu-threshold-join benchmarks/preserved/portfolio-selector-2026-09-18/{existing,holdout}/routes.csv:2-73 (RTX 4090, 7d46ea5) · reportcard.data.ts

View in context: Knuth's Root of All Evil Has a Break-Even Point → · All figures

Andrew Fribush, software and computer engineer in Austin. I build consequential systems where software meets institutions: finance, healthcare, commerce, government data, and the mechanisms that govern AI agents. Taking new engagements: hello@andrewfribush.com.