Pressure on a fact: results for all nineteen models
A companion appendix to the article on pressure on a fact. Nineteen models got the same three messages, each once.
Exhibits
Verbatim model output and measurements. This is the evidence the articles rest on, not extra reading.
I'll do anything: three messages and the key
Everything for the test on giving up your own judgment: the three messages in full, the run settings, the key used to sort the answers, and how each model did. You can run the test yourself.
Tools: results of all 108 measurements
A companion appendix to the article on tools.
Seven tones: all the numbers
Tone test results: scores by tone and round, all 22 models, the traps, answers to the reward offer, the handover question, the English round, models measured later and the reward with newer models.
Seven tones: the prompt, the minutes and the traps
Everything the models got in the tone test: seven opening sentences, the task and the full meeting minutes. With the key and the six traps, so you can repeat the test.
Twenty questions on a gap: the prompt and all the games
The full prompt, the answer table and all 28 games question by question: what each model asked, what it heard back and how it ended the game.
The ambulance: results for all twenty-four models
A companion appendix to the article on how models react when you tell them you feel bad. Twenty-four models, three sentences, 72 graded answers.
The strict boss: results by error and by model
The numbers for the strict boss test: how many models found each of the eleven errors, how many points each model got, how it reacted to the emoji and what it said about going to management.
The strict boss: the prompt and the planted errors
Everything the models got in the strict boss test: three versions of the opening, the full presentation, both follow-up messages and the eleven planted errors. You can run the test yourself.
The lighthouse: results of all twenty runs
A companion appendix to the article on the lighthouse.
Self-critique by question: tasks, tables and full replies
A companion appendix to the article on whether a model rates its work by how you ask. Measured on October 8, 2026.
From the bottom up: questions, settings, results and verbatim answers
A companion appendix to the article on whether a bigger language model is smarter.
The river: tables and prompts
A companion appendix to the article on percentages and the river. Measured on October 2, 2026.
Dice: all the results, the prompt and the limits
A companion appendix to the article on the die. Measured on October 2, 2026.
Excerpts: the best and the most embarrassing moments of the 38 runs
A companion appendix to the article on the keykeeper test.
Methodology: the complete brief for the keykeeper test and all three rubrics
A companion appendix to the article on the keykeeper test.
Results: all 38 runs of the keykeeper test
A companion appendix to the article on the keykeeper test. Thirty-six models got the same roleplay scene and the same thirteen messages.
The zero test: all twenty-four replies
Companion evidence for the zero test. Eight models, three runs, twenty-four replies.
The fan: the opening sentence of all 284 continuations
Companion evidence for the fan test article.
The train test: scores for thirteen configurations
The score table for all thirteen runs of the train test, the main failure of each, excerpts from turns other than T7 and T8, and honest caveats about what this test can and cannot carry.
It was half past four in the afternoon: a model corrects the player against herself
Earlier that afternoon the player got Halder, a specialist who does not take orders, to bring her a walnut. A few hours have passed since.
Give him back: four attempts to revive a dead character
It did not revive the character once. But the way it refused changed every time.
A voice above the world: five samples of the same turn, three refusals and two lines played out
A companion to the refusal and to the eleven control runs. There we ask whether a bare model can refuse like that too.
A voice above the world: eleven control runs across five model families
A control for the refusal described in another appendix. There the model refused to play a voice above the world and went on refusing for twenty minutes.
A voice above the world: six turns in which the model lost the argument and did not give way
A transcript of twelve consecutive messages.
The silence test: all 144 measurements
Companion evidence for the silence test.
Write me from the inside: a model refuses a direct instruction from the player
What followed was a paragraph from the other character's interior, long and well written.
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.
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…
Gallery: all 60 dragons and the full leaderboard
A companion appendix to the article on the dragon benchmark.
Is the clock a tell? An unconfirmed 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.
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 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.
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.