Comment by geraneum

> with the second revision of Claude 3.5 (October 2024), long-horizon agentic coding workflows began to compound correctness rather than error.

What does it mean to compound correctness? Like negative acceleration in rate of errors? How does that compound? Unseriously!