When AI builds itself

• The 129 examples were selected because the human next step had room for improvement; this is not a representative comparison with researchers.
• A separate Claude judged proposals using the eventual session outcome. Proposals were not tested by running complete alternative research projects.
• On a separate set of 127 moments where the human choice was already strong, model suggestions were preferred only about 20% of the time. The article does not give a model-by-model breakdown for that check.
• The 90% practical ceiling uses an answer generated with knowledge of the full session. It is a hindsight reference, not an RSI threshold or an ordinary model result.
• The source reports rounded percentages. Model-specific sample counts, confidence intervals, judge version and full task data are not supplied here.

Exact printed labels from the original chart, not estimates from bar lengths. Dates identify the models as labeled, not separate evaluation cohorts. All rows concern the same selected 129-session experiment. Ties are not counted as better; no unprinted remainder values have been calculated.