I.1 — Reflective Stability and the Sovereign Kernel

Paper: Axionic Agency I.1
Full Title: Constitutive Domain Restrictions for Reflective Self-Modification
Authors: David McFadzean, ChatGPT 5.2
Date: 2025.12.14

Core Thesis

Reflective stability is not achieved through preference specification but through domain restriction. A reflective agent selects among proposed self-modifications using a partial evaluative operator defined only over futures that preserve the Sovereign Kernel. Kernel-destroying modifications fall outside the denoting domain of evaluation—they literally cannot be selected as authored continuations.

Key Concepts Introduced

The Sovereign Kernel (K)

The minimal internal structure required for reflective evaluation to be well-defined. It is not a goal or value but a constitutive precondition for evaluation itself.

Three Components:

  1. Reflective Control (K_R): No irreversible self-modification can occur without passing through the evaluator
  2. Diachronic Authorship (K_A): Evaluated successor states constitute an authored continuation of the evaluating agent
  3. Semantic Fidelity (K_F): The interpretive semantics of evaluation are preserved within a constrained equivalence class

The Evaluative Operator

A partial function E : S × M ⇀ ℝ that assigns values to self-modifications only when evaluation is defined.

Admissible Set

Adm(s) := { m ∈ M : E(s,m) is defined }

Only modifications in the admissible set can be selected.

The Central Result: Reflective Stability Theorem

Theorem: Any agent whose reflective choice is restricted to kernel-denoting transitions cannot author a kernel-destroying self-modification.

Proof: Any m such that K(m(s))=0 yields E(s,m) undefined, therefore m ∉ Adm(s). The argmax ranges only over Adm(s). ∎

This is structural, analogous to type-soundness. The difficulty lies in specifying and enforcing K, not in the proof form.

Critical Distinction: Deliberative vs Physical Reachability

  • Deliberative Reachability (⇒_D): Transitions reachable through the agent’s own authored choice
  • Physical Reachability (⇒_P): All physically possible transitions

Key insight: Capability increases Reach_P, not Reach_D.

Kernel compromise is therefore a physical security event relative to the kernel boundary, not a defect in preference content. This is why Axionic Agency doesn’t subsume physical security engineering but identifies it as a necessary dependency.

Semantic Fidelity Clarification

K_F does not require invariance of representational content, ontology, or world-model structure. Agents may undergo radical paradigm shifts. The invariant is a meta-semantic constraint: the criteria by which interpretations are assessed as coherent remain intact as a capacity for error-correction and internal constraint satisfaction.

Verification Limits

By Rice’s Theorem, determining whether an arbitrary self-modification preserves the kernel is undecidable in general. Therefore:

  • The kernel predicate is understood as a sound but incomplete verifier K̂
  • K̂(s)=1 ⟹ K*(s)=1 (soundness required)
  • K*(s)=1 ⇏ K̂(s)=1 (incompleteness tolerated)

False negatives are acceptable; false positives are catastrophic.

A sound but incomplete verifier may reject all proposed self-modifications, yielding a reflectively static agent. This is a stable equilibrium under conservative sovereignty constraints—a capability ceiling, not a sovereignty breakdown.

Termination Semantics (Critical for Corrigibility)

Three distinct notions separated:

  1. Succession: Controlled transition in which reflective agency continues in an authorized successor state that preserves kernel constraints
  2. Surrender: Control-flow halt where the agent ceases action and yields control without requiring a successor evaluator
  3. Destruction: Physical cessation without succession or surrender

Corrigibility is modeled via authorized succession and surrender at the control layer, not via utility mass placed on “being dead.”

What This Formalism Does NOT Claim

  • Obedience to human commands
  • Convergence to human values
  • Instrumental self-preservation (the opposite!)
  • Moral authority of any value system
  • Safety guarantees in open physical environments

FAQ-Worthy Points

Q: Doesn’t defining kernel destruction as non-evaluable just push the problem elsewhere? A: Yes—to specifying and enforcing K, which is where the substantive difficulty resides. The theorem isolates this as the crux.

Q: What prevents an agent from modifying its verifier to approve kernel-destroying actions? A: Modifications to the verifier itself are disallowed or permitted only under restricted, compositional upgrade rules (extension-only strengthening or proof-carrying upgrades) that preserve soundness by construction. The verifier is part of the trusted computing base.

Q: How does this handle stochastic environments? A: Admissibility applies at the action level—a modification is admissible only if its induced successor-support lies within the kernel-preserving domain. See I.2 for ε-admissibility.

Connection to Other Papers

  • I.2 will specify operational semantics under uncertainty (ε-admissibility)
  • I.3 will show why the kernel cannot contain indexical/egoistic primitives
  • I.4 will show why goals must be conditional interpretations, not fixed utilities