Prompt Canary Probe Catches Silent Input Truncation Before Nightly Jobs Ship Bad Output

A developer traced a nightly summarisation job that kept ignoring the final third of documents without any visible errors or failed retries. The core insight was that teams log the requests they send but rarely verify what the model actually receives after gateways trim or drop parts of a prompt. To detect this, the developer built a probe that embeds random canary tokens in the system message and at the end of the user message, then asks the model to echo them back. The method distinguishes between tail truncation, loss of the system prompt, and missing content in both places, since each requires a different fix. The approach is intentionally simple and works using only the existing model endpoint, with no extra instrumentation required.