MB5 — Tiling
Tiling and ontology shift: can you trust a successor when the world-model underneath goals is rebuilt? Precise bet: full value-bundle and bearer transport through the ontology shift compose into successor safety.
What decision changes?
Before certifying a successor, ask whether the transport check spans the actual ontology change the successor introduces, not just the properties that were easy to measure.
In the field this is tiling and Vingean reflection, plus ontology identification (the diamond-maximizer worry). Can an agent trust a successor it cannot fully verify? Does a goal even survive when the world-model underneath it is rebuilt? Saying “the successor shares our values” is cheap if the ontology that interprets those values has shifted out from under them.
This project’s precise bet is MB5: full value-bundle transport plus bearer transport surviving an ontology shift are assumed to compose into successor safety . The check runs over the seven conserved properties , not over slogan agreement. That is a strengthening over “the successor says the right words,” but the composition itself is still assumed, not derived from something more basic. Closely related is MB10, which asks whether that green checklist can be forged.
Where agendas agree: MIRI tiling / Vingean reflection / ontology identification. Where they diverge: this project separates successor audit gaming as MB10; natural abstractions ease but do not discharge transport. Split: MB5 = transport holds; MB10 = successor certification checklist not forgeable.
Diagnostic evidence shows the practical gap: a relabeled or discontinuous successor identity can be invisible to light instrumentation and only show up once the audit splits the epoch and probes across it.
What would count as evidence?
Evidence would include successor-transition audits that detect relabeling, epoch shifts, or identity discontinuities that a narrower check would miss.