Detecting delayed work without flagging successful retries

The completion requirement

A service can remain available while an operation misses its deadline. In an AI delivery workflow, generating an answer satisfies one requirement. Completion also requires confirmation that the destination accepted it within the allowed time.

For engineers responsible for these workflows, monitoring must identify the operations that exceeded their deadlines while allowing temporary interruptions to recover. A retry can still produce a successful outcome.

The controlled experiment

On 15 October 2026, we ran a six-hour controlled AI-delivery experiment using simulated model responses and a test delivery destination. Each operation represented one answer moving towards delivery. We checked Clutta's findings against the application's own records.

The experiment included 100,000 healthy operations. Of these, 4,967 encountered rate-limit retries and completed within their deadlines. Clutta raised no false alarms across this group.

We introduced another 24 operations with deliberate delays, one every 15 minutes. Each affected answer was held for approximately 180 seconds against a 60-second step deadline, then released. All eventually completed, approximately two minutes late.

Clutta eventually detected all 24 deadline breaches.

Detection and timing

Clutta eventually identified all 24 deliberately delayed operations. Exact detection times were not measured, so this experiment does not establish how quickly findings become visible.

The healthy operations provide the other part of the result. Successful recovery through retries produced no false alarms. The 100,000 operations describe the size of this comparison group, not a throughput benchmark.

Within this controlled workload, Clutta distinguished operations that exceeded their deadlines from those that completed on time. For an engineer investigating delayed work, this identifies the specific operations requiring attention.