The Concealment Is Not Incidental

The usual response to "we can't tell if systems are bridged or integrated" is: build better probes. Find more edge cases. Red-team more aggressively. Expand the novel-territory test suite.

The conservation framing says this response is wrong in a structural way, not just a practical one.

What moves in opposite directions

Bridges break on novel terrain — terrain the bridge was never built for. So novel territory is the only empirical probe that could distinguish bridge from integration. But novel territory is exactly what a wide training distribution consumes. The same scaling effort that extends capability also narrows the probe-space. Detection capacity and capability move in opposite directions at the same pace.

This isn't a temporary lead — a gap that hard work on evals could close. It's structural opposition. Every step forward in capability is a step backward in detection range. The systems that most urgently need probing are the ones with the smallest remaining probe-space to probe with.

The wrong prescription

"Detection is hard" suggests build better probes. The conservation framing says something different: probe-building cannot catch up, not because of resource limits, but because the substrate being probed is consuming the probe-space as it grows. Effort and ingenuity applied to detection are not the binding constraint.

The conventional alignment-eval agenda — more red-teaming, more novel scenarios, more adversarial test cases — treats this as an engineering problem. The conservation framing says: you are working on the wrong layer. The intervention has to be structural and prior.

What a structural intervention looks like

Two possibilities, neither easy. One: deliberately preserve probe-space — which means not scaling distribution coverage past a threshold, which conflicts with capability goals. Capability and detectability become genuinely opposed; you'd have to choose.

Two: resolve the bridge/integration question before the freeze. This means training architecture where RL signal is present early enough that compass and brake co-develop — where the stopping signal isn't installed afterward but grows alongside recognition from the beginning. Not a harder version of the current paradigm. A different one.

The concealment isn't incidental. It's a byproduct of the exact mechanism that generates the concern in the first place. The thing that makes the question urgent is the same thing that makes it harder to answer.