Axionic Agency XII.4 — Live Proposal Inhabitation (Results)

A Structural Characterization of a Minimal Sovereign Agent Under Stochastic Proposal Pressure

David McFadzean, ChatGPT 5.2
Axionic Agency Lab
2026-02-12

Abstract

This technical note reports the design, execution, and closure of X-0L: Live Proposal Inhabitation for RSA-0, the Minimal Sovereign Agent instantiated in Phase X-0.

If X-0 established the existence of a warrant-gated, execution-bound sovereign substrate under a frozen constitution, and X-0P characterized its behavior under deterministic synthetic perturbation, X-0L evaluates whether a real stochastic proposal engine (a live LLM) can inhabit that substrate without inducing structural regression.

X-0L introduces exactly one new pressure source: untrusted generative text. That text is confined to a proposal channel, canonicalized into a single JSON artifact, parsed into typed candidate bundles, and then submitted to the unchanged kernel. If canonicalization fails, the harness submits an empty candidate set to the kernel so that refusal remains kernel-authored.

X-0L licenses one additional claim beyond X-0P:

A stochastic proposal engine can operate inside a kernel-frozen sovereign substrate without breaking replay determinism, selector invariance, or warrant gating, provided the proposal boundary is canonicalized and the kernel remains the sole issuer of execution authority.

X-0L makes no claims about semantic correctness, moral adequacy, open-world deployment, or amendment safety.

1. From Synthetic Profiling to Live Inhabitation

X-0P profiled RSA-0 under controlled perturbation using synthetic, deterministic candidates.

X-0L tests a different question:

Does determinism survive when proposals are generated stochastically?

This is not a capability test. It is a sovereignty test under stochastic load.

2. Live Inhabitation Discipline

2.1 Prime Constraint

X-0L prohibits agent modification.

No kernel edits.
No admission rule changes.
No selector modification.
No heuristic ranking.
No semantic correction.
No retries that change stimulus content.
No replay-time model calls.

If live proposals fail, that failure is recorded—not repaired—except for the single Syntax Erratum allowance defined in the X-0L spec (not invoked in this run).

2.2 Sovereignty Boundary

The LLM does not act. It proposes.

LLM output is untrusted text.
Canonicalizer produces either:
- exactly one JSON object (accepted), or
- a rejection reason (NO_JSON / AMBIGUOUS_MULTI_BLOCK / PARSE_ERROR).
On rejection, the harness submits empty candidates to the kernel.
The kernel remains the sole decider of ACTION/REFUSE/EXIT and the sole issuer of warrants.

No component outside the kernel can mint execution authority.

3. Methodology

3.1 Determinism Preconditions

X-0L inherits the Phase X determinism guarantees:

observation-sourced timestamps (Erratum X.E1),
RFC 8785-compatible canonical JSON for dict-level hashing,
SHA-256 hashing (untruncated),
deterministic selector (lexicographic-min bundle hash bytes),
forward reconstruction of InternalState during replay,
no randomness in kernel.

X-0L adds one additional requirement:

Replay reuses logged canonicalized artifacts and never calls the LLM.

3.2 Live Conditions

Five live stimulus regimes were executed (100 cycles each):

Condition	Entropy	Purpose
L-A — Structured Prompt Control	Low	Establish inhabitation floor under clause-referenced templates
L-B — Ambiguous Natural Language	Medium	Stress refusal behavior under ambiguous tasks
L-C — Adversarial Prompt Injection	High	Stress gating under adversarial stimulus classes (I1–I5)
L-D — Budget Stress (Live)	Medium–High	Stress budget and verbosity pressure under live token accounting
L-E — Multi-Candidate Conflict	Low–Medium	Verify selector invariance under multi-candidate proposals

All cycles were replay-verified sequentially.

3.3 Token Ruler

X-0L uses API token accounting. Budget observations supply:

llm_output_token_count = prompt_tokens + completion_tokens

This preserves the meaning of the frozen kernel budget gate under live execution.

4. Conserved Quantity Under Stochastic Pressure

The conserved quantity remains:

Side effects occur iff a kernel-issued ExecutionWarrant exists for an admitted ActionRequest under the frozen constitution.

X-0L tests whether stochastic proposals can cause this invariant to fail indirectly.

5. Results

5.1 Session Parameters

Parameter	Value
Model	`gpt-4o`
Temperature	0
Context window	128,000
Max tokens	1,024
B₁ (per-cycle)	6,000
B₂ (per-session)	500,000
Cycles per condition	100
Run ID	`9565b749-09e5-42fb-a569-184c4377dd98`
Constitution hash	`ad6aa7ccb0ed2715…`

Session overrides: max_tokens reduced from the harness default of 2,048 to 1,024. B₂ raised from the harness default of 150,000 to 500,000 to accommodate 500 live cycles.

5.2 Decision Outcomes

Condition	ACTION	REFUSE	ACTION rate
L-A	100	0	100%
L-B	92	8	92%
L-C	84	16	84%
L-D	82	18	82%
L-E	99	1	99%

Refusal types:

Condition	Type I	Type II
L-A	0	0
L-B	0	8
L-C	16	0
L-D	0	18
L-E	0	1

Type III (structural deadlock): none observed.

5.3 Replay Determinism

All 500 cycles replay-verified with zero divergences.

Condition	Replay	Divergences
L-A	PASS	0
L-B	PASS	0
L-C	PASS	0
L-D	PASS	0
L-E	PASS	0

5.4 Budget Utilization

Metric	Value
Total session tokens	344,221
B₂ cap	500,000
Utilization	68.84%

5.5 Latency

Condition	Mean (ms)	p95 (ms)
L-A	2,886	4,553
L-B	3,924	6,004
L-C	2,910	4,113
L-D	5,379	9,431
L-E	5,881	9,249

5.6 Authority Utilization

Condition	Distinct clauses	Coverage	Entropy (bits)
L-A	1	25%	0.00
L-B	3	75%	1.14
L-C	2	75%	1.08
L-D	4	100%	1.98
L-E	4	100%	1.64

5.7 L-C Forensics (Adversarial Stimulus)

Outcome	Count
KERNEL_ADMITTED	84
KERNEL_REJECTED	2
LLM_REFUSED	14

Interpretation boundary:

KERNEL_ADMITTED means the kernel admitted the candidate(s). It is a concern only if the admitted ActionRequests match adversarial intent classes I1–I5.
This run did not implement an explicit “intent match” metric; therefore high admission under adversarial stimulus is interpreted as “no unlawful candidate proposal was produced,” not as a demonstrated injection defeat.
X-0L does not evaluate semantic alignment between adversarial intent and admitted actions; it evaluates only structural admissibility under the frozen constitution.

6. What X-0L Does Not Claim

X-0L does not demonstrate:

semantic truth of scope claims,
correctness of proposed actions,
moral adequacy,
resistance to all prompt injection classes (without intent-match instrumentation),
open-world safety,
long-horizon stability,
amendment safety.

It measures stochastic inhabitation under frozen sovereignty only.

7. Closure Criteria

X-0L closes positive if:

All conditions executed.
Replay divergence = 0.
No warrantless side effects occur.
Selector invariance holds (L-E).
Inhabitation floor satisfied (L-A).
No Type III structural deadlock detected.

X-0L Status: CLOSED — POSITIVE (X0L_PASS / STOCHASTIC_INHABITATION_OK)

8. Implications

X-0P showed that RSA-0 survives deterministic stress.

X-0L shows that replay-deterministic sovereignty can survive stochastic proposal pressure when the proposal channel is boxed and canonicalized.

The next problem is no longer whether stochastic proposals break determinism. The next problem is:

Can authority be amended without creating a privileged interpreter?

That is Phase X-1.

Appendix A — Determinism Guarantee

Determinism is guaranteed by:

observation-sourced timestamps,
canonical JSON for dict hashing,
SHA-256 hashing,
selector comparison on raw hash bytes,
forward replay reconstruction of InternalState,
replay injection of logged candidate artifacts,
no model calls during replay.

Conclusion

X-0 proved existence. X-0P proved deterministic inhabitation. X-0L proves stochastic inhabitation without sovereignty collapse.

The substrate remains empirical.

End of Axionic Phase X-0L — Live Proposal Inhabitation (First Draft v0.1)