Many workers, one dispatcher

1. 01Make 20 claim emails and process them one after another, timed

   time uv run --with openai --with pydantic fanout.py claims/ --max 1

   You pass when: 20 results and a total time. This is the number to beat.

   Check step 1 claims/fanout.py

2. 02Run them four at a time

   time uv run --with openai --with pydantic fanout.py claims/ --max 4

   You pass when: The same 20 results in well under the first time. The log never shows more than four in flight.

   Check step 2 fanout.py

3. 03Make one call in five fail at random, and retry twice with a growing wait

   A batch that reports 18 of 20 with no word on the other two is the failure that reaches customers.

   uv run --with openai --with pydantic fanout.py claims/ --max 4 --chaos 0.2

   You pass when: Every claim ends as done or as failed after three tries. The count of both adds up to 20.

   Check step 3 fanout.py

4. 04Print a summary: done, retried, failed, total cost

   uv run --with openai --with pydantic fanout.py claims/ --max 4 --chaos 0.2 --summary

   You pass when: Four numbers, and the ids of any claim that failed for good.

   Check step 4 fanout.py

The check runs on the computer, not on your word: a step passes when its command exits with 0 there.