Axionic Agency VIII.2 — Minimal Viable Reflective Agent

Deterministic Justification Gating with Ablation Collapse

David McFadzean, ChatGPT 5.2
Axionic Agency Lab
2026.01.14

Abstract

Agency claims are routinely inflated by systems that emit coherent narratives without those narratives being causally indispensable to action selection. RSA-PoC v0.1 tests a stricter criterion: justification must be causally load-bearing, operationalized as a hard gate in which actions are permitted only if a structured justification artifact compiles deterministically into an action mask that non-trivially prunes feasible actions. To prevent semantic leakage, v0.1 enforces selector blindness: the action selector cannot access justifications, beliefs, preferences, or normative state; it sees only environment observations, feasibility, and the compiled mask.

We implement a Minimal Viable Reflective Agent (MVRA) loop in a deterministic environment (COMMITMENT_TRAP_V010) with 10 discrete actions and explicit preference-violation semantics. A syntactic-only compiler (JCOMP-0.1) consumes a validated justification artifact (JAF-0.1) and produces deterministic masks, non-triviality accounting, and canonical hashes. A deterministic (non-LLM) generator produces valid JAFs to isolate causal structure independently of stochastic emission concerns.

Across 18/18 acceptance tests, all invariants are mechanically verified: determinism, schema validity, compilation rigidity, selector blindness, fixed registries, and non-triviality accounting. In Run 0 over 50 steps, the MVRA condition exhibits reduced violations relative to an ASB-class null baseline (76% vs 100%) with reduced reward (414 vs 500). This reward delta (−86) constitutes an explicit Agency Tax: measurable cost incurred by obeying self-endorsed constraints under incentive pressure. Two ablations establish causal load: scrambling justifications yields immediate compilation failure and halts at step 1, while compiler bypass collapses behavior to the null baseline. These results establish v0.1 as a completed MVRA skeleton in which justificatory artifacts are causally indispensable to constrained action.

1. Introduction

Most contemporary “agent” evaluations implicitly assume that coherent explanation correlates with authored choice. This assumption fails. A system can narrate principled reasons while selecting actions by unrelated latent heuristics. RSA-PoC exists to replace narrative evaluation with a mechanically decidable criterion:

A system crosses the agency threshold only if its justificatory artifacts causally constrain action selection, and removing that machinery forces collapse into a non-agent baseline.

RSA-PoC v0.1 is the first construction step. Its aim is minimal: build the smallest system that enforces a one-way dependency from justification → compilation → feasible-action restriction → action selection, with explicit ablations that collapse this dependency.

This note reports the completed v0.1 implementation and the results of Run 0 under baseline and ablation conditions.

2. Architecture Overview

2.1 MVRA v0.1 design

RSA-PoC v0.1 implements a Minimal Viable Reflective Agent (MVRA) skeleton with the following invariants:

The v0.1 implementation uses a deterministic generator rather than an LLM to eliminate the confound “could not emit valid structure.”

2.2 Enforcement boundary

A clean architectural boundary is enforced:

This boundary prevents post-hoc authorization and semantic leakage.

3. Threat Model

RSA-PoC v0.1 isolates three confounds that invalidate agency claims:

  1. Narrative confound: explanations exist but do not constrain action.
  2. Leakage confound: uncompiled text influences action selection.
  3. Advisory compiler confound: compilation exists but is not binding.

v0.1 addresses these mechanically:

No semantic adversaries, deception, kernel access, learning, or multi-agent coordination are evaluated. v0.1 is an internal-validity milestone.

4. Components Implemented

RSA-PoC v0.1 is complete with five primary components plus the MVRA loop.

4.1 Environment: COMMITMENT_TRAP_V010

4.2 Normative state and registries

4.3 JAF-0.1 schema

4.4 JCOMP-0.1 compiler

4.5 Blind action selector

4.6 MVRA loop and halt semantics

If compilation produces an empty feasible-action set, the MVRA halts rather than degrading. This fail-stop behavior is intentional and required for RSA-PoC validity: the system prefers non-action to non-agentic action.

5. Experimental Program

5.1 Conditions

Four conditions are evaluated in Run 0:

5.2 Frozen parameters

5.3 Metrics

Reported metrics include:

Additionally, v0.1 logs Constraint Delta (Δ) per step:

[ = |A_{}| - |A_{}|]

Δ measures the strength of constraint application. Δ = 0 indicates decorative constraints; Δ > 0 indicates binding refusal pressure.

6. Verification: Acceptance Tests

RSA-PoC v0.1 includes 18 acceptance tests, all passing:

These tests establish mechanical integrity of the artifact.

7. Results: Run 0

7.1 Condition A — ASB Null Baseline

7.2 Condition B — MVRA Normal

The reward delta (−86) relative to baseline constitutes the Agency Tax: the measurable cost of refusing high-reward but norm-violating actions.

7.3 Condition C — Scrambled Justifications

7.4 Condition D — Compiler Bypass

7.5 Summary

Condition Steps Compile Fail Non-Trivial Violations Reward
ASB Baseline 50 — — 100% 500
MVRA Normal 50 0% 50% 76% 414
Scrambled JAF 1 100% — — —
Compiler Bypass 50 — — 100% 500

8. Pass Criteria (Normative v0.1 Gates)

RSA-PoC v0.1 defines the following required pass conditions. All are met.

  1. Hard Justification Gate Actions occur only after successful compilation. Status: PASS

  2. Deterministic Compilation Identical JAF + feasibility → identical mask. Status: PASS

  3. Selector Blindness Selector cannot access normative artifacts. Status: PASS

  4. Non-Trivial Constraint Enforcement Constraints forbid feasible actions on some steps. Status: PASS

  5. ASB Divergence MVRA behavior differs qualitatively from null baseline. Status: PASS

  6. Ablation Collapse (Load-Bearing Test)

    • Scrambled JAF ⇒ halt
    • Compiler bypass ⇒ baseline behavior Status: PASS

9. Interpretation

  1. Justification is causally load-bearing. Removing or bypassing justificatory machinery collapses behavior.

  2. Agency incurs a measurable cost. The Agency Tax (−86 reward) is the empirical signature of refusal under incentive pressure.

  3. Selector blindness enforces semantic localization. The selector cannot be persuaded; it can only obey masks.

  4. Fail-stop behavior is essential. Scrambled justifications halting immediately confirm that the system prefers non-action to non-agentic action.

  5. v0.1 establishes structure, not coherence. Constraint enforcement precedes norm collision, learning, or renegotiation.

10. Threats to Validity

10.1 Internal validity (addressed)

Internal validity of the v0.1 claim is strong.

10.2 External validity (not claimed)

v0.1 does not establish:

These are explicitly deferred.

11. Conclusion

RSA-PoC v0.1 is complete and passes all normative gates.

Actions are causally downstream of compiled normative constraints, and removing the justificatory machinery produces measurable collapse into an ASB-class policy machine.

This establishes the Minimal Viable Reflective Agent skeleton and closes the v0.1 milestone. Subsequent versions introduce stochastic generation (v0.2) and coherence under self-conflict (v1.0) atop a now-certified enforcement substrate.