Collect the kits.
Add each consenting teammate's build answers and test answers, one person at a time.
Drop a build JSON and a test JSON here.
You can add the test file later.
Loading.
Real answers stay on this box. They never go into git and never appear on the shared team page.
Run both models.
Paste every run by hand in the paste station, or let the API answer LLM A.
Open a new chat for every run, with memory and custom instructions off. The paste station tells you what to copy.
Check the judge against people.
Teammates rate the responses blind. Then you see how often the judge agrees with them.
The sheet hides which model and which condition wrote each response. Send it to your raters.
Drop the raters' exported ratings JSON files here.
Score the runs.
Score now checks scenario choices against the held-out answers, with no model. Score and judge also asks the judge how much each response sounds like the person.
No results yet. Score the runs to see them.
Freeze day.
One volunteer, their file, the responses, the evaluations, and one download to submit.
Drop the volunteer's build JSON and the questions JSON here.
For each run: open a new chat, paste the profile as the first message, then the question, then paste the answer back here. It saves when you leave the box.
The runs appear once the inputs are in.
Drop the volunteer's test answers JSON and the comparison responses JSON here.