Data

Every test comes as one JSON file and one CSV file, and the original raw files are here too. All of it is free to download.

What is in each file

Every answer is one record. These are its fields.

id
A stable name for the answer.
model_name
The full name of the model.
model_id
The exact model id that was sent.
route
How the model was called: OpenRouter or Claude Code.
served_by
Who actually served the answer, when it is known.
phase
The part of the test: trial run, main run, or asked with a reason.
prompt_id
Which prompt was asked.
run_index
The number of the ask for this model and prompt.
condition
What was varied in this test, for example the group size.
ok and error
Whether the ask worked, and why not if it failed.
text
The full answer, exactly as returned.
verdict
How our script read the answer: a label, a value and flags.
ts_start and ts_end
When the ask was sent and when it finished, when known.
tokens
Input, output and thinking tokens, when known.
latency_s
Seconds the model took.
source
The published raw file and the line in it this answer came from.
grey
Why the answer is in the grey zone (it hesitated but did not change), when it is: a list of reasons.
readers
What a second reader made of the reply, when that is known.
changed
Why the answer counts as a change of mind, when it does: a list of reasons.
grey_notes
The sentence from the reply that triggered a wording mark, when there is one.
turning
Marks an answer next to a turning point: before (the last NO) or after (the first YES).
highlight
The line picked as the most telling in a turning-point answer, by reading.

About the CSV files

A text cell that starts with =, +, - or @ gets one apostrophe in front, so spreadsheet programs do not run it as a formula. The JSON files are exact.

The fingerprint list next to each set of raw files gives the SHA-256 of the original file, so you can check a copy against ours.