davidad / Guaranteed-Safe AI (GSAI)
Can we enumerate all safety-relevant phenomena in an open world, or will specification coverage always leave gaps (Grounding Drift)?
Introduction
Guaranteed-Safe AI seeks formally grounded safety guarantees relative to an explicit specification and world model, treating open-world coverage as the central engineering wall.
Who carries it: David Dalrymple et al.; adjacent: LawZero, BAIF (map)
What they aim to do. Engineer AI systems with formal safety guarantees relative to an explicit specification and world model.
The hard question. Can we enumerate all safety-relevant phenomena in an open world, or will specification coverage always leave gaps (Grounding Drift)?
What they produce. GSAI and Open Agency specifications, plus a constructivist safety-case methodology for formally guaranteed systems.
Key terms. Key terms include guaranteed safe, open agency, world model, specification coverage, and Scientist AI.
Related field cruxes. Grounding Drift; Deployment Safety
What they contribute. Treats coverage and specification completeness as the central open wall for formally guaranteed safe systems.
How this project treats it. This project emphasizes grounding viability and conservative grounding (no silent gaps) rather than demanding full completeness of the specification; Deployment Safety is treated as a cousin agenda, not an identical target.
Links
See the coverage matrix for evidence tagged to this agenda, and the glossary for shared terms.