It passed once. Would it pass again?

I want to check whether my AI agent’s results hold up when run again, using CheckAgain (checkagain.cruciblescience.com). Please:
1. Find the cases I already test (an eval set or tests in this project), or ask me for the 10 to 20 cases I try by hand after a change.
2. Write a small script in this project that runs each case 5 times through my agent, the way my code already calls it, and decides pass or fail with the check I already use. If there isn’t one, agree a plain rule with me first; don’t use a model as the judge unless I ask.
3. Take a setup label from a flag (--setup before / --setup after) and append one row per run to results.csv with the columns task, setup, run, passed, ended, cost_usd (ended: completed, timeout or error; cost only if my client records it).
4. Before running anything, tell me how many agent calls it will make and roughly what that costs, and wait for my yes.
Put only task names and outcomes in the file: never prompts, outputs, conversation text, customer data or keys. Don’t upload anything; I’ll drop the file into CheckAgain myself.

Turn my existing agent results into a CSV for CheckAgain (checkagain.cruciblescience.com). The results are in: [where your results are].
Write one row per run with the columns task, setup, run, passed (pass or fail), and, if the data has them, ended (completed if the run stopped by itself, otherwise what stopped it) and cost_usd (as recorded, never estimated). Keep each task’s name the same across runs, number repeats 1, 2, 3…, and use the model or version as the setup.
Copy results exactly as recorded: don’t invent, fix or drop runs. Put only labels and outcomes in the file: never prompts, outputs, conversation text, customer data or keys. Save it as results.csv and tell me where it is.