Reading the primary source, publishing the number the summaries leave out.

Index

Every claim traces to a paper, a filing, a release note, or code opened by hand. One note a day; a longer piece when the material earns it.

• 2026·11·03 0.016% of the weights, six times the maths score Astra fixed Qwen3.5-0.8B's reasoning loop by hand-editing six parts in ten thousand. The size of the edit, not the score, is the finding. ai systems · 1,000 words
• 2026·11·03 DeepSeek's Router doesn't do the work because 47 lines of prompt forbid it Source code, not the paper: 47 of 148 prompt lines are prohibitions, plus three arithmetic constraints that do the real enforcing. ai systems · source code · 1,100 words
• 2026·11·03 DeepSeek's Super-agent costs 3.8× more per run. That's the point. The token column nobody quoted: 6.1M → 23.2M cache reads for a shorter tail. A prompt fixes p50; only architecture fixes p95. ai systems · 1,300 words