To ground a study on a different corpus, clone it — editing this set would redefine the comparison its existing runs belong to.
Session
Time limitmin · blank = none
Blinding
Blinded
Blinded is the whole blinding surface. Clear-session, rewind, and example-prompt controls are always available to enrollees and need no setup.
Equal weights average to a plain arithmetic mean. Scores run 0–100.
Rewrite the enrollee's prompt before answeringenable_prompt_enhancer
Have a person approve the rewrite firsthuman_review_enabled
Reserved for a later version. Human review applies to ad-hoc use, not to a controlled study, because a per-enrollee edit is not a uniform condition.
Remember the conversation within a chat sessionenable_chat_history
How much is remembered is a platform setting, not a study setting — it is capped by the smallest context window among the models in the set-up, so every arm of the comparison gets the same. Enrollees can rewind or clear a session themselves.
One artifact holding both the questioning and the rewrite phrasings. Required once this study attaches context documents or declares required details — the assistant would otherwise have nothing to reply with. Choose the style per run at launch, to measure which gets people to an answerable request faster.
Required Details
which market or instrument the question is aboutrequired
the time period of interestrequired
Studying Socratic mode? This list decides what you are comparing. Fill it in to compare two ways of repairing a request; leave it empty to compare asking against not asking at all.
Clear as the run proceeds
Cleared as each round finishes. Leanest storage, and “See all” still works live during a round. Default.