manage-experimentslisted
Install: claude install-skill whilehq/whileai-sdk
# Manage experiments
You know the repo; the person reading the page does not have your context.
They read for two reasons: to learn what a paper or an idea does when
tried, and to find a behavior and fix it fast at work. Both want one
look: what changed, did it move, why, what it taught, and can I run it
again. On 2026-09-20 real agents posted `dapo-lr5e-05-s17-30st` nineteen
times, no note on any run, no picture, a `0.75` on a page that counts
points, and no question on six agents of seven. Every page was correct
and said nothing. Every step below is in `check.py`, which runs offline
in under a second; its setup defines `MODEL`, `OPEN_MODEL`, `TOOLS`,
`POLICY`, `PROMPT_POLICY`, `DATES_LINE`, `PROMPT_DATES`, `ASKS`, `NOISE`,
`STEPS`, `CHECKPOINT_EVERY`, `score_on`, `score_checkpoint`,
`train_reward`, `worst_of` (20 rows, failures first) and `fake`.
Names follow [platform/naming](https://docs.while.ai/platform/naming);
that part is easy.
## 1. The question, before the first run
One agent per product, one question at a time, at the top of the page.
```python
tracked = track("checkout-support", model=MODEL, transport=fake) # drop transport= for real
tracked.experiment(
question="Does telling the prompt today's date cut wrong refunds on old orders?",
hypothesis="Most wrong refunds are orders past 30 days; a dated rule removes them.",
method="Same asks, same judge. Harness: two prompts on one model. Training: SFT on traces.",
measure="refunds_when_eligible on