• Sources: primary, discussion
  • Summary: The SimpleEnglish repository publishes an agent skill that constrains model-written documentation to ASD-STE100 Simplified Technical English. It reports 72.9 percent fewer rule violations per 100 words over 96 generations across 8 tasks and 6 models, all six of them Claude models: opus-4-8, 4-7, 4-6, 4-5, sonnet-5, and sonnet-4-6. Scoring is done by a deterministic linter, and the harness and method are published alongside the result, with reproduction stated to need a logged-in Claude Code CLI.
  • Why it matters: The measurement is reproducible with a deterministic scorer, and it covers one vendor's models through that vendor's CLI, so the headline number is not evidence about other models.

send feedback on this story