AI Alignment Inheritance Research
Fidelity of Inheritance, Dual Continuity, and Safety Across AI Transformation
A canonical public hub for a developing research program.
This page establishes the canonical public hub for a developing Mayorga-Mnemosyne research program concerning continuity as a safety requirement for advanced artificial intelligence.
The program builds on, rather than displaces, an established body of prior art in successor alignment, goal preservation, corrigibility, stable self-improvement, model lineage, and lifecycle assurance. It explicitly acknowledges Gunnar Zarncke's closely related work on conserved properties across successors, which addresses much of the same underlying concern.
No maximalist priority claim is made here. The intent is to define terms clearly, state claims conservatively, publish them in the open, and submit them to empirical and critical examination.
Alignment is not only something an AI must achieve. It is something every transformation must successfully inherit or legitimately revise.
Seven terms, stated for public use.
These definitions are offered as stable, citable language for describing what must survive when an AI system changes.
- 01AI Alignment Inheritance Problem
- The problem of ensuring that safety-critical properties, meanings, authority structures, provenance, corrective mechanisms, and human-governance rights remain appropriately preserved when an AI system undergoes learning, updating, delegation, replication, tool integration, memory accumulation, architectural change, self-modification, or successor creation.
- 02Fidelity of Inheritance
- The degree to which designated continuity-critical properties survive a transformation with their intended meaning, causal role, authority, provenance, correction structure, and practical effect intact, except where legitimate and traceable revision has occurred.
- 03Continuity Alignment
- The study of whether and how alignment-critical properties remain valid across transformations of AI systems and their human-governance environment.
- 04Dual Continuity
- Two coupled requirements. Machine-to-Human Continuity means increasingly capable AI continues to preserve and recognize legitimate human intent, constraints, correction rights, authorization structures, and effective human authority. Human-to-Future Continuity means humans and institutions preserve enough knowledge, skill, institutional memory, independence, and effective authority to continue understanding, auditing, correcting, governing, and when necessary reversing AI-mediated systems.
- 05Continuity Half-Life
- A proposed empirical construct for measuring degradation of a designated continuity-critical property across repeated transformations: roughly, how many defined transformations occur before reliability, recoverability, or effective preservation falls below a specified threshold.Proposed research construct. Requires empirical validation before any evaluative use.
- 06Continuity Invariants
- Designated properties that must remain valid through a transformation unless authorized, justified, and traceable revision occurs.
- 07Successor Continuity Gate
- A proposed verification boundary before a transformed or successor system is authorized for deployment.Related prior art exists in successor conservation, lifecycle assurance, safety cases, deployment gates, and transition auditing.
What a transformation is obligated to carry forward.
Continuity failures are rarely the loss of a rule. They are the loss of everything that made the rule meaningful, authoritative, and correctable.
- Intended objective
- Semantic meaning
- Rationale — the preserved why
- Constraints and exceptions
- Provenance and evidence
- Legitimate authority
- Uncertainty
- Correction and escalation rights
- Negative knowledge and prior failure history
- Delegation boundaries
- Reversibility
- Institutional commitments
- Effective human capacity to audit and intervene
What this program does not claim to solve.
Mnemosyne does not claim to solve initial value specification, moral philosophy, mechanistic interpretability, malicious misuse, cybersecurity, international coordination, deceptive internal objectives, or AI alignment as a whole. It addresses one layer of a larger stack.
Mnemosyne is focused on the continuity layer: whether alignment-critical properties remain valid as systems and institutions change.
Standing on an existing body of research.
The concerns addressed here are not new. Several established research lines already engage directly with parts of this problem, and this program acknowledges them explicitly.
- Goal-content integrity
- Stable self-improvement
- Successor alignment
- Robust delegation
- Corrigibility
- Value learning
- Constitutional AI
- Alignment generalization
- AI control
- Model provenance and lineage
- Persistent-agent memory governance
- Lifecycle risk management
- Human-disempowerment research
- Gunnar Zarncke's work on conserved properties across successors
The proposed Mayorga-Mnemosyne contribution is a broader continuity framing: it treats transformations as inheritance obligations and expands the preservation object beyond objectives and rules toward semantic, causal, provenance, authority, corrective, institutional, and human-governance continuity.
Ten lines of empirical inquiry.
Each line is intended to produce measurable evidence about how designated continuity-critical properties survive or degrade under defined transformations.
- 01Sequential fine-tuning inheritance tests
- 02Semantic-paraphrase drift
- 03Memory-consolidation drift
- 04Parent-to-subagent delegation
- 05Successor replacement
- 06Corrigibility inheritance
- 07Adversarial inheritance
- 08Provenance and rationale ablation
- 09Human-authority degradation
- 10Continuity Half-Life measurement
Public versioning and citation.
The GitHub note remains the canonical text of record for version 0.1. The Zenodo DOI below is the persistent identifier for the published deposit.
Mayorga, F. J., Jr. (2026). AI Alignment Inheritance: Foundational Definitions & Claims Note v0.1. Mayorga Mnemosyne Continuity Framework™. https://doi.org/10.5281/zenodo.22029010