D4 — Retry storms

The problem

A timeout fires, the client retries, the provider bills both calls. Retry loops without idempotency guards or backoff can pay for the same request three, five, ten times — and the bill shows nothing but normal-looking calls.

Detection

Calls with identical prompt fingerprints (hash identity where available) on the same model are clustered inside a tight anchor window. A cluster of n > 1 identical calls in the window is n − 1 paid duplicates.

Savings estimate

wasted = (n − 1) × mean cost of the cluster's calls

Duplicate detection uses only priced rows — if some calls in a cluster hit an unpriced model, they are not imputed into the estimate.

Worked example (golden fixture)

5 identical gpt-5.4-mini calls, 20 seconds apart, 500 prompt / 200 completion tokens each:

per call = (500 × $0.75 + 200 × $4.50) / 1e6 = $0.001275
wasted   = 4 × $0.001275                     = $0.0051
monthly (×10)                                = $0.0510

Severity for D4 follows cluster size rather than dollars — a 10-deep retry cluster is an incident-in-waiting even when the model is cheap.

Known limitations

Without request_id or prompt hashes in your export, identity is inferred from token-count fingerprints — legitimate identical calls (polling the same prompt on purpose) can cluster. The report's evidence rows show you the timestamps so you can tell a storm from a schedule.