My AI Lost to Simple Rules. So I Built BenchSloth Around Evidence.
This is a submission for the Hacktoberfest Open-Source AI Challenge: Week 3 : Slow and Steady. The most useful result in my AI benchmark was the one telling me to give the model less authority. My first lab-note assistant recognized a wait—and then made it harder to use. Its model-driven policy offered just 3 of 13 eligible waits . Guarded rules offered 10 . The clever part was getting in the way…
This submission for the Hacktoberfest Open-Source AI Challenge focuses on the concept of "Slow and Steady." The author discovered that limiting the model's authority led to better results. The AI policy provided only three eligible waits out of 13 options, while guarded rules offered ten. The author realized that giving the AI too much control was unnecessary and instead built an agent that managed attention around evidence.
BenchSloth is the result of this idea. It provides a plan for attention rather than just an experiment. The tool includes three main parts: Source, Attention, and Intelligence. The Source keeps the original words and occurrence IDs intact, while the Attention shows fixed intervals and review needs in source order. The Intelligence section allows for inspecting evidence selected for a question.
When using BenchSloth, users can ask which earlier wait a note describes, open quoted passages, inspect the actual model request, record corrections, confirm intervals, and return to the next instruction. The agent records the path between these actions. However, it does not claim to have watched or completed the experiment.
The demo starts with the clear wait (A), which includes a Gemma response, measured duration, and verification trace. The uncertain note (B) maintains the distinction between two identical waits, and the longer workflow (C) consists of thirteen lines across two sections. By returning to the initial wait (A) and confirming its ten-second demonstration, users can breathe easier knowing that the clock and source evidence remain visible.
The source material is preserved in the three parts of BenchSloth, and the agent's path is deterministic. If the source lacks an exact duration or relevant target, the tool cannot grant a timer. BenchSloth successfully recovered 22 out of 22 reference links and 16 out of 18 eligible waits in synthetic cases.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
