AI Coding Benchmark Exposes How Compressed Summaries Outlive Their Own Corrections
A local AI coding benchmark originally conducted in June was re-tested on 4 July using qwen3-coder-30b via Ollama, which scored a perfect 8/8 in just 2.1 minutes compared to 7.5/8 in 44.1 minutes for the smaller qwen3.5-9b model. The re-test initially appeared to overturn the June verdict, but closer reading revealed the original findings had been misrepresented through compression — the phrase 'iterative coding not viable' had been stripped of its architectural scope and specific conditions. June's analysis had already identified five structural failure modes and noted that a proper harness could address at least two of them, which is precisely what the July setup provided. The core problem was not that the June verdict was wrong, but that a condensed version of it entered the project's working memory as a headline claim and went unchecked for two weeks. The episode illustrates how accurate technical findings can be undermined when nuanced conclusions are reduced to broad, unqualified statements during documentation handoffs.


