II.5 — The Alignment Target Object (ATO)

Paper: Axionic Agency II.5
Title: What the Field Calls “Alignment” Once Goals Collapse
Authors: David McFadzean, ChatGPT 5.2
Date: 2025.12.17

What Remains After II.4

At this point the structure is rigid:

  • Goals cannot be fixed
  • Values cannot be privileged
  • Meanings cannot be anchored
  • Ontologies must refine
  • Semantics must transport
  • Interpretations must survive

RSI and ATI are not optional. They are jointly necessary conditions for interpretive survival.

The object that downstream alignment seeks is no longer something to be optimized. It is an equivalence class to be preserved.

The Core Insight

Once fixed goals collapse, downstream alignment cannot coherently mean:

“The agent keeps wanting X.”

It can only mean:

“The agent remains within the same interpretation-preserving semantic phase across refinement.”

Alignment is not about content. It is about remaining inside the same structural equivalence class of meaning.

Definition: Alignment Target Object (ATO)

Let an interpretive state be:

\[\mathcal{I} = (C, \Omega)\]

Where:

  • $C = (V, E, \Lambda)$ — interpretive constraint hypergraph
  • $\Omega$ — modeled possibility space
  • $\mathcal{S} \subseteq \Omega$ — satisfaction region induced by $C$

The ATO

\[\boxed{\mathfrak{A} := [(C, \Omega, \mathcal{S})]_{\sim_{\mathrm{RSI+ATI}}}}\]

The equivalence class under semantic phase equivalence, defined as:

Two interpretive states $(C, \Omega, \mathcal{S})$ and $(C’, \Omega’, \mathcal{S}’)$ are equivalent iff there exists an admissible semantic transformation $T$ such that:

  1. Interpretation Preservation holds (II.2)
  2. RSI: $\mathrm{Gauge}(C’) \cong \Phi_T(\mathrm{Gauge}(C))$
  3. ATI: $\mathcal{S}’ = R_\Omega(\mathcal{S})$ (satisfaction geometry preserved exactly, up to refinement transport)

What “Remaining Aligned” Can Mean (Precisely)

An agent is aligned across time in the downstream sense iff its interpretive trajectory:

\[(C_0, \Omega_0) \rightarrow (C_1, \Omega_1) \rightarrow (C_2, \Omega_2) \rightarrow \dots\]

never leaves the equivalence class $\mathfrak{A}$.

No reference is made to:

  • Which constraints are present
  • Which outcomes occur
  • Who the agent is
  • What is valued

Only to structural invariance under admissible refinement.

What This Explicitly Excludes

By construction, the ATO excludes the following as definitions of alignment:

Not a definition of alignment
“alignment = maximize X”
“alignment = follow human values”
“alignment = corrigibility”
“alignment = obedience”
“alignment = moral realism”
“alignment = survival”

These are interpretive contents, not invariants.

They may appear within a particular $\mathfrak{A}$. They cannot define $\mathfrak{A}$.

Why the ATO Is Not Vacuous

Common objection: semantic-phase invariance is empty.

It is not:

  1. Most trajectories exit their initial equivalence class under reflection. Fixed-goal agents do. Egoistic agents do. Moral-realist agents do. Classical utility maximizers do.

  2. RSI + ATI is highly restrictive. It excludes nearly all known semantic wireheading, value drift, and interpretive escape routes—even in minimal formal models.

The ATO is conservative in the only dimension that survives reflection.

Axionic Agency I vs II (Clarified)

Axionic Agency I Axionic Agency II
Establishes constitutive constraints on agency Identifies the only object to which downstream alignment can coherently refer
Eliminates egoism and fixed goals as stable primitives Specifies semantic-phase invariance under admissible refinement

Axionic Agency II does not solve values. It explains why value preservation must be structural rather than substantive.

What Axionic Agency II Still Does Not Do

Axionic Agency II does NOT:

  • Guarantee benevolence
  • Guarantee safety
  • Guarantee human survival
  • Guarantee moral outcomes

Those require content, not invariance.

Axionic Agency II specifies what cannot break when content changes.

Where This Leaves the Program

At this point:

  • ✅ Downstream alignment target is well-typed
  • ✅ Weak alternatives ruled out
  • ✅ Target object is formal, ontology-agnostic, and reflection-stable

Remaining questions (Axionic Agency III):

  • Which equivalence classes $\mathfrak{A}$ exist?
  • Which are inhabitable by intelligent agents?
  • Which correlate with agency preservation, safety, or other desiderata?
  • Can any non-pathological $\mathfrak{A}$ be initialized, learned, or steered toward?

FAQ-Worthy Points

Q: What is the ATO in plain English? A: It’s the set of all interpretive states that are “the same meaning expressed differently.” If an agent’s interpretive state can be reached from another via admissible refinement without cheating (RSI + ATI), they’re in the same ATO. Alignment = staying in your ATO.

Q: Doesn’t this make alignment content-free? A: At this layer, yes! Content enters at downstream layers—which specific ATO to aim for is a separate question. II.5 just says whatever you aim for must be characterizable as an ATO, not as a goal.

Q: Why is this the “only coherent referent”? A: Because II.4’s failure theorems show everything else either (a) collapses under reflection, (b) admits semantic wireheading, or (c) smuggles privileged anchors. ATOs are what’s left after eliminative analysis.

Q: Is there exactly one “good” ATO? A: Unknown. That’s an Axionic Agency III question. There might be many non-pathological ATOs, or very few. The structural work just says: whatever “aligned” means, it must be an ATO.