AI Anonymous / Practical field guide / November 12, 2026

Wrong inbox?

Two AI summary jobs start in one order and finish in another. Pair the results by position and both summaries land on the wrong tickets. Reproduce the swap in an offline Python lab.

In this field note · 6 sections
  1. Reproduce the swap
  2. Keep the ticket ID with the future
  3. Make it your own
  4. Apply the mapping in a batch
  5. Read the Python references
  6. Run the lab or share the field note

Imagine a support dashboard that drafts short summaries for two tickets. You submit a password reset first and a missing receipt second. The receipt job finishes first. Both summaries read fine, but the dashboard silently swaps them if it assigns the first finished result to the first submitted ticket.

What we ran. Windows CPython 3.13.15, two threads, and two handwritten summaries. A thread event forces the receipt job to finish first so the swap is reproducible without guessing how long to sleep.
AI Anonymous poster reading “Wrong inbox?” A wax-sealed envelope lies diagonally across two oak mail trays with brass label holders. JABSystems.io appears below.
Original AI-generated artwork of a wax-sealed envelope straddling two oak mail trays.

01 / Reproduce

Reproduce the swap

Extract the practice ZIP and open PowerShell or a terminal in the extracted folder. Use Python 3.10 or later. The exercise uses the standard library and runs offline.

python batch.py
python check.py

submit() returns a Future for each call. The first worker waits on an event. With two workers, the collector can see the second future finish and then release the first. The script prints submission order, completion order, and both mappings. The check script verifies six assertions, including the error case below.

submission_ids = [ticket["id"] for ticket in tickets]
outcomes = [outcome for _, outcome in completed]
by_position = dict(zip(submission_ids, outcomes, strict=True))

The input IDs are T-201, T-202. The completion order is T-202, T-201. Both lists have two items, so zip(..., strict=True) accepts them even though their positions refer to different tickets.

02 / Inspect

Keep the ticket ID with the future

At submission, store the original ticket ID under its future. as_completed() yields each finished or cancelled future. The lab saves each completed future with its outcome and then looks up that exact future to identify the ticket.

future_to_id = {
    pool.submit(summarize, first, gate): first["id"],
    pool.submit(summarize, second): second["id"],
}

# Each (future, outcome) pair comes from the completion loop.
by_future = {future_to_id[future]: outcome for future, outcome in completed}

Read the two mappings from the same run.

T-201 by position  Receipt missing.
T-201 by future    Reset link expired.

T-202 by position  Reset link expired.
T-202 by future    Receipt missing.

The ID comes from the input and stays in the application. The worker only returns a summary. If the dashboard needs submission order for display, look up the finished outcomes by ID in that order after collection.

03 / Make it your own

Give one ticket a failed job

  1. Copy tickets.json to my-tickets.json.
  2. Change the second ticket’s summary to JSON null. The fixture worker raises a ValueError and the collector records it as a readable error.
  3. Run the modified file. Check the ID that gets the error under each mapping.
python batch.py my-tickets.json

The positional mapping blames T-201. The future mapping records the error under T-202, the job that actually failed. For your own batch, start with two fictional rows from its input format and assert the association for both a success and a failure.

Boundaries

Apply the mapping in a batch

The practice script expects two rows with unique IDs. Its event only controls the demonstration. In your collector, assign a unique ID per submitted job and preserve it when you handle an error from Future.result(). Review the output separately for summary quality before any write. This lab tested local result association and the authored failure case. Model calls, dashboard updates, provider timeouts, retries, and performance remain untested.

Primary sources / Checked November 12, 2026

Read the Python references

Downloads / Practice and social files

Run the lab or share the field note

Explore more AI Anonymous field notes →

Share this post.

Found this useful? Pass it along. Download the artwork, add your own take, and tag @jabsystems.io so I can see your post.