The Preservation Benchmark: Testing Functional Continuity Across Substrate Transfer

This preprint introduces PreservationBench, a benchmark framework for measuring preservation relevant functional continuity after a source agent is copied, transplanted, adapted, or relearned in a target substrate. The benchmark is not a consciousness test and does not decide personal identity. It asks which source specific functions remain active after transfer, and which apparent recoveries are better explained by relearning, behavioral mimicry, or proxy optimization. As an initial validation tier, the paper implements PreservationBench AST v0 in a toy neural agent inspired by Attention Schema Theory. A source agent is trained in a small grid world task with attention control, self report, and source specific identity probes. Selected source modules are transferred into a target architecture under matched controls: frozen copied state, source alignment repair, copied attention, source facing attention and control bridges, behavior only distillation, and frozen random controls. In an expanded core replication, 30 independent source seeds were attempted and 22 passed source validation. Frozen copied schema and self model state preserved source specific identity probe behavior but did not recouple report or control. Long source alignment repair improved report but did not restore source like task control. A source facing attention bridge recovered report behavior without source like action agreement. A control bridge drove reward and goal attention mass above source levels but remained weak on goals found and source action agreement, indicating proxy optimization rather than preserved control. Copying source attention yielded the strongest combined profile, with identity probe preservation, high self report, near source goals found, and high source action agreement, but this condition narrows the substrate gap.