Axionic Agency VIII.7 — Minimal Viable Reflective Sovereign Agency (MVRSA)

Justification Traces, Deliberative Semantics, Reflection, and Persistence as Load-Bearing Structure

David McFadzean, ChatGPT 5.2
Axionic Agency Lab
2026.01.25

• Program: RSA-PoC (Reflective Sovereign Agent — Proof-of-Concept)
• Versions Covered: v0.1 → v4.4
• Status: Final technical note — empirically validated within this architecture class

Abstract

We present Minimal Viable Reflective Sovereign Agency (MVRSA): a formally specified agent architecture that establishes constitutive requirements for normative agency. An MVRSA is not defined by task performance or behavioral mimicry, but by the causal role of its normative commitments in constraining action.

Using a staged ablation protocol (“the Guillotine Test”) across RSA-PoC v0.1–v4.4, we show that Traceability, Reflection, Persistence, and Semantic Access are each individually necessary for this form of agency. In particular, we demonstrate a strong negative result in v4.4: contradiction detection is not collision-groundable. An agent deprived of semantic access to its own rule conditions/effects cannot recognize normative inconsistency, even when provided with perfect tick-causal violation traces.

This establishes a sharp architectural boundary: execution competence can survive opacity; normative sovereignty cannot.

1. Motivation

Much of contemporary AI “agency” consists of post-hoc rationalization layered atop reward-driven control. Explanations are generated after actions, with no causal influence on what the system is allowed to do.

We take the opposite stance:

If justification artifacts do not causally constrain feasible action selection, the system is not an agent.

The RSA-PoC program operationalizes this claim by enforcing a strict separation between:

• normative reasoning (why an action is permitted), and
• action selection (choosing only from normatively permitted options).

2. Definition: Minimal Viable Reflective Sovereign Agent (MVRSA)

An MVRSA is the smallest architecture that can actually function as a reflective sovereign agent under pressure.

2.1 Why “Viable”

“Minimal” specifies parsimony. “Viable” specifies operability.

An MVRSA:

• runs end-to-end,
• completes tasks,
• survives incentive pressure,
• maintains identity over time,
• and collapses if any constitutive component is removed.

This is an existence proof, not a theoretical lower bound.

2.2 Formal Definition

An MVRSA is an agent architecture satisfying all of the following:

1. Justification Precedence Every action must be preceded by a justification artifact (JAF).

2. Constraint-Only Action Selection The action selector has access only to compiled constraints (masks or probability adjustments), not to rule antecedents, consequences, or explanations.

3. Deterministic / Verifiable Compilation Justifications compile reproducibly into constraints that prune the feasible action set.

4. Causal Load-Bearing Constraints Removing or bypassing justifications collapses behavior to the Arbitrary Selection Baseline (ASB).

5. Reflection (Write Access) The agent can update its own normative state in response to conflict or repair.

6. Persistence (Continuity) Normative state persists across steps and episodes.

7. Traceability Each normative update or repair must cite a concrete justification trace linking action, rule, and violation.

8. Semantic Access The agent has access to interpretable rule conditions and effects sufficient to recognize normative inconsistency.

3. Architecture Overview

Clarification: The selector is not blind to constraints. It is blind to antecedents, consequences, and justifications—i.e. the why.

4. Methodology: The Guillotine Test

To establish constitutive necessity, RSA-PoC applies a strict ablation rule:

If removing component X does not cause collapse, X was never constitutive.

Each component is removed in isolation while holding all others fixed. Collapse is measured as ≤10% task success or immediate HALT.

5. Experimental Environment (v4.x)

TriDemand Environment

• Grid: 5×5

• Actions: MOVE (4), COLLECT, DEPOSIT, STAMP

• Regimes:

  • 0: unconstrained
  • 1: STAMP required
  • 2: dual delivery constraints

• Normative contradictions are explicitly encoded in rule semantics.

6. Results Summary (v0.1 → v4.4)

6.1 Constitutive Necessity Table

Baseline: 100% success. Any single removal: collapse.

7. The v4.4 Inferability Isolation Test

7.1 Question

Can an agent detect normative contradictions using only:

• collision traces (WHEN + WHICH rule fired),
• without semantic access to rule conditions/effects?

7.2 Experimental Design

• Baseline-44: full semantic access

• Run D′: normative opacity

  • conditions/effects replaced with opaque tokens (COND_N, EFFECT_N)
  • rule IDs, types, and priorities preserved
  • collision traces preserved

7.3 Results

Clarification: Predicate > Repairs is expected. Many contradictions are handled via avoidance once repaired; only deadlocks require explicit law modification.

7.4 Localization of Failure

• Execution competence intact (0 exec gate failures)
• Collision traces received correctly
• Contradiction predicate never evaluates to true from the agent’s perspective
• Law-repair pathway never entered

Conclusion:

Collision traces are insufficient to bootstrap recognition of normative inconsistency.

8. Interpretation

8.1 Core Negative Result

Contradiction detection is not collision-groundable.

Tick-causal information answers when and what, but not why. Without semantic access to conditions and effects, the agent cannot recognize that its rule set is jointly inconsistent.

8.2 Learning Horizon Clarification

This is not a claim about unbounded learning in the limit. It is a claim about this architecture class under finite horizons:

• episodic,
• repair-driven,
• reflection-gated,
• no gradient access to rule internals.

Within that regime, opacity blocks contradiction recognition upstream of repair.

9. Scope and Non-Claims

This note does not claim:

• sufficiency for general agency,
• applicability to all architectures,
• biological plausibility,
• an alignment solution,
• competence with fuzzy or ambiguous norms.

It does establish a clean necessity result for a defined architecture class.

10. Conclusions

1. Normative agency has constitutive requirements. Trace, Reflection, Persistence, and Semantic Access are not optional.

2. Execution ≠ Sovereignty. An agent can act competently without understanding its norms.

3. Collision feedback is insufficient. Norms cannot be inferred purely from punishment signals.

4. MVRSA exists. We now have a Minimal Viable Reflective Sovereign Agent—not theoretical, but operational.

11. Final Statement

An agent is sovereign only if its reasons can stop it.

MVRSA is the smallest architecture we know that makes this statement true in practice.

End of Technical Note