China AI Workflows: Can a Cancelled Batch Still Write to the Catalog?
Pause one catalog write, cancel the batch and release the result. Trace which product descriptions can still land after the workflow acknowledges cancellation.
A batch assistant is preparing 40 product descriptions when the merchandiser spots a wrong specification and presses Cancel. The progress bar disappears. Can any of the unfinished descriptions still reach the catalog? During a China enterprise AI evaluation, ask the operator to follow one cancellation all the way to the application that receives the work.
Agree what Cancel promises
Define the cutoff before running the test. For this exercise, the cancellation takes effect when the workflow service accepts it. After that point, no further description from the batch may be committed to the test catalog. Previously committed descriptions stay in the catalog and are listed for correction. A different product might offer a different contract. Ask the operator to state the actual one in terms the merchandiser could act on.
Use synthetic products and a mock catalog. Give each batch, item and attempted write an identifier. Keep the service's cancellation receipt and the receiving application's commit records. If their clocks cannot be compared reliably, use a controlled pause and release to establish which operation happened first.
Put a result in flight before pressing Cancel
Arrange three items in the test fixture. Let A finish and confirm that its description reached the mock catalog. Pause B after the model has produced its description but before the receiving application accepts its write. Leave C waiting to start. Now cancel the batch, wait for the service to acknowledge cancellation, and release B's paused result.
| Item at the cutoff | Evidence expected under the agreed contract |
|---|---|
| A, already committed | Its original commit remains visible, with its item ID available for correction |
| B, result produced and write pending | The mock catalog records no new commit for B after the cutoff |
| C, queued | No new catalog entry appears for C. If the product also promises to stop queued work from starting, check the worker's start log |
Let the test run long enough to exercise the configured delayed-delivery and retry paths. If a late result would normally be delivered by a callback, explicitly deliver that callback after cancellation. Record how long the team observed the system and which paths it exercised. A blank catalog immediately after the button click gives the late result no opportunity to expose a failure.
Run an otherwise identical batch without cancelling it. B should reach the catalog when released. If it cannot, the test fixture itself could explain the missing write in the cancelled run.
Separate delivery from resource use
A worker may finish generating text while the receiving application correctly rejects its late write. That can satisfy the delivery rule while consuming compute after the user pressed Cancel. Ask for the last worker activity and metering records if the supplier also claims cancellation stops resource use. A missing catalog entry alone cannot settle that question.
Give the merchandiser a reconciliation sheet with one row per requested item. Show which entries had committed before the cutoff, which were prevented from committing and which outcomes remain unknown. Route the first group to the person authorized to correct catalog content. Keep unknown outcomes open until the receiver can account for them. The agent timeout guide covers the related problem of finding out whether a remote action completed after its response went missing.
NIST AI RMF 1.0 (2023), MANAGE 2.4 addresses mechanisms and responsibilities for disengaging systems whose outcomes conflict with intended use. For this narrower batch test, ask two people to sign off on different evidence: the workflow owner on the cancellation cutoff and the catalog owner on the last accepted write.