An empirical field note testing whether increasing `reasoning_effort` from medium to high improves LLM review quality. Using GLM-5.2 and Codex on synthetic documents with planted contradictions, the author found that higher reasoning effort narrowed GLM-5.2's detection breadth (2.0 to 1.4 issue types per run) while broadening Codex's (2.0 to 2.6). The key takeaway: reasoning effort is not a universal 'IQ dial' — its effect is model-specific and text-specific. The piece also models good epistemic practice, showing how a single compelling observation dissolved under repetition and cross-text testing.
Table of contents
The first story was almost too satisfyingFive runs made the certainty evaporateA different text removed the problem altogetherThe better question was breadthWhat I would do differently in a real review workflowShare this post