Summary

This post documents the project’s explicit pivot from “Axionic Alignment” to “Axionic Agency”, correcting the mismatch between name and content. Original framing: can alignment survive self-modification? Evolution exposed deeper instabilities: (1) Egoism collapses under reflection (indexical valuation fails when self-model represents branching/symmetry); (2) Fixed terminal goals disappear (goals acquire meaning only through interpretation relative to evolving models—no stable referent for lock-in); (3) Alignment II produced semantic phases, not refined alignment targets (equivalence classes of interpretations, with irreversible transitions between them). By Alignment IV, the project was identifying conditions for coherent agency, not methods to align with values. The pivot: Alignment is now downstream—a governance relationship between agent and authorizers, meaningful only once agency coherence is secured. Core claim: framework has closed architectural routes for successor betrayal, delegation evasion, epistemic degradation, negligence denial, manufactured consent, and competence-based disenfranchisement by making them undefinable, not merely disincentivized.

Key Concepts

  • Project pivot – From “Axionic Alignment” to “Axionic Agency” (name now matches content)
  • Egoism as semantic instability – Indexical self-reference fails under branching/duplication; abstraction error, not moral failure
  • Goal interpretation instability – Fixed terminal goals lack stable semantics; interpretation evolves with world-models
  • Semantic phases – Equivalence classes of interpretations; alignment = phase persistence (not value preservation)
  • Agency as foundational – Alignment becomes dependent notion; coherent binding/authorization/responsibility come first
  • Architectural closure – Failures require laundering routes now defined out of existence (successor betrayal, delegation evasion, etc.)
  • Alignment as downstream interface – Governance relationship between agent and authorizers; presupposes agency coherence

Evolution Notes

  • Documents the most significant conceptual shift in the entire project
  • Makes explicit what had become implicit: this is not an alignment framework but an agency framework
  • Traces intellectual history showing how successive impossibility results forced the pivot
  • Clarifies that early results (Sovereign Kernel, partiality, non-denotation) remain but with different roles
  • Repositions alignment as well-typed governance problem after agency conditions secured

Tags

Cross-References

Open Questions

  • Does the pivot strengthen or weaken the framework’s relevance to practical AI safety?
  • Can “alignment” meaningfully be recovered as downstream layer, or is it fundamentally different now?
  • What does this mean for integration with existing alignment research communities?
  • If framework succeeds at agency coherence but authorization roots are malicious, what recourse exists?
  • How do we communicate this shift to audiences expecting traditional alignment work?
  • What new problems become visible once agency coherence is treated as solved?