Position paper · 20261017
Pseudo-SOTA on Domain Benchmarks: Is It the Tool's Scaling Dividend or the Quality of the Harness Mechanism?
HAVEN: on a domain benchmark the model is fixed, so what is really compared is the harness. A fixed-model protocol with three controlled conditions that separates harness mechanism quality from the tool dividend.
Cite
scalingX. “Pseudo-SOTA on Domain Benchmarks: Is It the Tool's Scaling Dividend or the Quality of the Harness Mechanism?.” RsiingSI, 2026.
scalingX. “Pseudo-SOTA on Domain Benchmarks: Is It the Tool's Scaling Dividend or the Quality of the Harness Mechanism?.” RsiingSI, 2026.
https://rsiing-notes.pages.dev/HAVEN-Harness-Attribution-EN.pdf