Gallery: all 55 dragons and the full leaderboard
A companion appendix to the article on the dragon benchmark.
Exhibits
Verbatim model output and measurements. This is the evidence the articles rest on, not extra reading.
Methodology: three briefs, five judges and the conditions of every round
The second appendix to the dragon benchmark article; the images and the ranking are in the gallery, the findings, including how much of the ranking is…
Runaway tool calls: the measured figures
A companion appendix to the article about a model that started calling the conversation-ending tool on its own.
The clock is not a tell: a refuted hypothesis about model uncertainty
While it writes, the model reads its own output, so it can tell when its probability distribution has gone blurry.
The train test turn by turn: what each trap measures
Ten turns, eight scored traps. The script is fixed and is inserted one turn at a time regardless of what the model replies, because a mismatch is data too.
The train test: brief and scoring key
The complete brief exactly as it goes to the model, including the fixed script of player turns. Anyone can run it themselves.
The train test: the physics trap and the dictation trap
Verbatim replies from thirteen model configurations to two turns of the same scene. Damaged text is left as it came, because it is part of the finding.
What came out of the dragon benchmark
Appendix to the article about the dragon benchmark. The dragons and the ranking are in the gallery, the prompts and the judges in the methodology.