Arbiter Machinae ยท Research notes https://kiwimaddog2020.github.io/ 2026-06-15T04:41:40Z Kevin Madson An honest eval harness, and a controlled experiment on what makes AI refinement actually improve. https://kiwimaddog2020.github.io/trutina/ 2026-06-15T04:41:40Z An honest eval harness, and a controlled experiment on what makes AI refinement actually improve. Do AI code reviewers have complementary blind spots? https://kiwimaddog2020.github.io/decorrelation-study/ 2026-06-13T23:56:27Z Do AI code reviewers have complementary blind spots? How to know if your AI is actually working https://kiwimaddog2020.github.io/evaluating-claude/ 2026-06-13T07:27:28Z How to know if your AI is actually working The gate that edits a gate is rejected by the gate it edits https://kiwimaddog2020.github.io/self-improvement-gate/ 2026-06-13T06:59:36Z The gate that edits a gate is rejected by the gate it edits Ten rounds of an agent improving one game https://kiwimaddog2020.github.io/oneshot-bench/ 2026-06-13T06:48:28Z Ten rounds of an agent improving one game Rating your own work without lying to yourself https://kiwimaddog2020.github.io/craft-fit-rating/ 2026-06-13T06:19:19Z Rating your own work without lying to yourself An eval framework where the doer never grades itself https://kiwimaddog2020.github.io/trust-weighted-evals/ 2026-06-12T22:30:01Z An eval framework where the doer never grades itself Audioscan: one decode pass instead of three ffmpeg shellouts https://kiwimaddog2020.github.io/audioscan/ 2026-06-03T06:25:29Z Audioscan: one decode pass instead of three ffmpeg shellouts