MB2 — Bundle Identifiability

Behavioral and internal traces are assumed to identify stable value-bundle geometry well enough to license claims of bundle alignment.

What decision changes?

Before claiming two systems share values, ask whether the measurement distinguishes shared bundle geometry from shared surface behavior under a narrow training distribution.

Inverse reinforcement learning is underdetermined: many reward functions fit the same behavior, and separating values from irrationality is an open problem the CIRL/assistance-game literature inherits directly. The book’s response is to drop scalar reward for value-bundle geometry — but that move only helps if bundle geometry is itself identifiable from evidence.

MB2 is the assumption that it is: that behavioral and internal traces pin down stable bundle geometry closely enough to call two systems bundle-aligned. It does not resolve the pointing problem so much as relocate it into a named, checkable measurement claim.

The toy-simulation diagnostics show this bridge failing under light instrumentation specifically — bundle drift can be invisible to correction handles that are not tracing bearer welfare, which is exactly the failure mode this bridge would need to rule out to be trustworthy at low measurement cost.

What would count as evidence?

Evidence would include held-out distribution shift tests where bundle-equivalent systems keep agreeing and non-equivalent systems come apart.