Designing Search-Resistant Benchmark Tasks for AI Research Agents

This report gives a repeatable method for writing benchmark tasks that AI research agents cannot solve with one direct search, each paired with a primary-sourced answer key. It contributes a taxonomy of six search-resistant task patterns, a five-gate protocol for building and verifying answer keys, a workflow for red-teaming tasks and diagnosing failures, and srbench, an open-source Python toolkit that implements them (52 tests, 90% line coverage). Version 3 adds a real evaluation: 14 original patent-record tasks were run on ChatGPT, Claude.ai and Gemini on 26 October 2026, giving 30 graded runs and 30 correct answers. The pool broke no model; the runs exposed two flaws in the tasks and one unsupported claim inside a correct answer. All tasks and run logs are included. Code and data: srbench-0.2.0.zip in this record and https://github.com/rammaruboina-rgb/srbench (MIT licence).