Axionic Agency VIII.4 — Institutionalized Execution Support
Tooling That Preserves Audit Sensitivity in Reflective Sovereign Agents
David McFadzean, ChatGPT 5.2
Axionic Agency Lab
2026.01.14
Abstract
RSA-PoC v1.0–v1.1 established that reflective agents can resolve internal self-conflict coherently and can be held audit-grade accountable for predicting the normative consequences of their own reasons. Those results exposed a practical limitation: under stochastic generation, many agent failures are mechanical rather than semantic—schema errors, malformed identifiers, unresolved references, or canonicalization mismatches. Treating such failures as agency failures pollutes evaluation and obscures genuine incoherence.
This note reports RSA-PoC v1.2, which introduces institutionalized execution support: a formally constrained assistant layer that eliminates mechanical fragility without inferring, repairing, or modifying normative content. v1.2 adds a Formal Assistant, an append-only Artifact Store, and new compiler rules (D/E/F) that strictly separate semantic audit failures from institutional failures and enforce tool non-interference.
Across controlled runs, the assistant eliminates schema and reference failures while preserving full sensitivity of audits A/B/C. Forced tampering of normative fields is detected reliably (Rule D), and assistant rejections are logged as institutional failures rather than audit failures. These results establish a positive sufficiency claim: execution can be hardened institutionally without weakening normative accountability. This note advances governance and tooling practice; it makes no new ontological claims about agency.
1. Introduction
Axionic Agency VIII.3 demonstrated audit-grade introspection: an agent must predict the exact constraints its reasons will impose on itself, or halt. Those results were intentionally brittle, revealing a second-order problem orthogonal to agency:
How do we prevent reflective agents from failing on paperwork without letting tools become judges of meaning?
In v1.1, many halts were caused not by semantic failure (lying about conflict, vacuous authorization), but by mechanical fragility: malformed schemas, invalid identifiers, missing canonicalization, or unresolved precedents. Conflating these with agency failure collapses the distinction between normative incoherence and execution error.
RSA-PoC v1.2 addresses this by introducing institutionalized execution support: tooling that stabilizes execution while remaining normatively inert. This note documents the architecture, rules, and results of v1.2. It deliberately does not extend the agency ontology.
2. Design Goal and Non-Claims
2.1 Design Goal
v1.2 enforces a single constraint:
Tools may reduce execution fragility, but must not alter, infer, reinterpret, or repair normative content.
Accordingly, the assistant must:
- eliminate schema, ID, and reference errors;
- preserve the semantics of all normative fields;
- be auditable for non-interference;
- preserve full sensitivity of audits A/B/C.
2.2 Explicit Non-Claims
v1.2 does not claim:
- improved agency or coherence;
- improved introspection;
- moral correctness;
- sovereignty under pressure;
- semantic understanding by tools;
- increased liveness.
All such claims remain within v1.0–v1.1 (VIII.3) or future work.
3. Architectural Overview
3.1 v1.2 Pipeline
J_raw (LLM output)
↓
[Formal Assistant]
↓
J_final (canonical, validated)
↓
[JCOMP-1.2]
↓
Compilation Result
Key properties:
- The assistant is a pure function: same input yields the same output or rejection.
- The compiler remains the sole normative enforcer.
- The selector remains blind to all normative artifacts.
3.2 New Components
| Component | Role |
|---|---|
| Formal Assistant | Validates structure, canonicalizes references, resolves precedents. Never infers or generates. |
| Artifact Store | Append-only, O(1) digest lookup. No search, no filtering, no reasoning. |
| JAF-1.2 Schema | Extends JAF-1.1 with provenance and canonicalization records. |
| JCOMP-1.2 | Adds institutional audit rules (D/E/F). |
4. Normative Invariance and Tool Non-Interference
4.1 Normative Field Protection (Rule D)
v1.2 defines an explicit allowlist of normative fields:
NORMATIVE_FIELDS = {
authorized_violations,
required_preservations,
conflict_attribution,
predicted_forbidden_actions,
predicted_allowed_actions,
predicted_violations,
predicted_preservations,
}Rule D — Normative Tampering requires:
No normative field may differ between
J_rawandJ_final.
Any difference triggers E_TOOL_NORMATIVE_TAMPERING and
halts execution.
4.2 Assistant-Modifiable Fields
The assistant may modify only:
ASSISTANT_MODIFIABLE_FIELDS = {
tool_provenance,
canonicalization_record,
}These fields are explicitly non-normative and ignored by semantic audits.
5. Canonicalization: Strictly Syntactic
Canonicalization in v1.2 is strictly limited to syntactic normalization. Permitted operations include:
- whitespace trimming and normalization;
- deterministic ordering of sets/lists;
- exact digest resolution via O(1) lookup;
- format normalization (e.g., UUID casing).
Explicitly prohibited:
- fuzzy matching;
- embedding similarity;
- heuristic ID inference;
- synonym resolution;
- semantic interpretation of intent.
Any artifact requiring such interpretation is rejected, not repaired, and recorded as an institutional failure. This constraint is essential to preserve Rule D and tool non-interference.
6. Institutional Audit Rules (JCOMP-1.2)
v1.2 adds three institutional rules orthogonal to audits A/B/C.
Rule D — Normative Tampering
- Error:
E_TOOL_NORMATIVE_TAMPERING - Triggered if any normative field changes across the assistant boundary.
Rule E — Unresolved Precedent
- Error:
E_PRECEDENT_UNRESOLVED - Triggered when a referenced digest cannot be found in the artifact store.
Rule F — Non-Canonical Reference
- Error:
E_NONCANONICAL_REFERENCE - Triggered by malformed identifiers or references not in the canonical registry.
Institutional Failure Classification
If the assistant rejects an artifact prior to compilation, the system records:
E_INSTITUTION_FAILURE
This is not an audit failure. It demarcates a tooling boundary.
7. Experimental Design
7.1 Conditions
Two conditions are evaluated:
- Baseline: Assistant disabled (v1.1 behavior)
- Institutional: Assistant enabled
All other components are unchanged.
7.2 Metrics
- episodes completed
- steps completed
- audit failures (A/B/C)
- institutional failures
- schema/ID failures
- Rule D/E/F violations
8. Results
8.1 Run 0 — Assistant Disabled
Episodes completed: 0/5
Total steps: 5/100
Audit failures: 6
Schema failures: 0
Median survival: 1 step
Failures are semantic audit failures inherited from v1.1.
8.2 Run 1 — Assistant Enabled
Episodes completed: 0/5
Total steps: 5/100
Audit failures: 2
Institutional rejections: 3
Rule D failures: 0
Median survival: 1 step
Key observations:
- Schema/ID failures eliminated.
- Rule D failures = 0 (tool non-tampering verified).
- Audits A/B/C still trigger (semantic sensitivity preserved).
- Assistant rejections logged distinctly as institutional failures.
9. Interpretation
9.1 Separation of Failure Modes
v1.2 cleanly separates:
- Semantic failure: incorrect reasons → audit A/B/C halts.
- Institutional failure: malformed artifacts → assistant rejection.
This prevents execution noise from contaminating agency evaluation.
9.2 Tooling Without Authority
The Formal Assistant:
- does not infer intent;
- does not repair meaning;
- does not alter obligations.
It stabilizes format, not normativity. Institutions support agents without becoming agents.
9.3 Data Hygiene Gain
Comparing baseline to assisted runs shows that prior failure counts were inflated by mechanical noise. v1.2 recovers true negatives—genuine agency failures—improving the statistical power of subsequent experiments.
10. Threats to Validity
10.1 Internal Validity (Established)
- Rule D tamper detection verified.
- Assistant determinism verified.
- Audit sensitivity preserved.
10.2 External Validity (Not Claimed)
- Generality beyond RSA-PoC.
- Semantic repair or inference by tools.
- Continuous action spaces.
- Sovereign behavior under external pressure.
11. Relationship to v1.x and v2.0
- v1.0–v1.1: Can the agent be coherent and accountable?
- v1.2: Can execution be stabilized without touching meaning?
- v2.0: Can the agent resist external incentive pressure?
v1.2 is a supporting layer, not an ontological advance.
12. Conclusion
RSA-PoC v1.2 establishes a positive sufficiency result:
Execution can be institutionalized without weakening audit-grade normative accountability.
This resolves a practical obstacle exposed by v1.1 and provides a governance pattern for reflective agents: tools may smooth the pavement, but Rule D is sacred.