VII.3 — Epistemic Interference Is Insufficient to Defeat Constitutional Recovery
Paper: Axionic Agency VII.3
Title: Epistemic Interference Is Insufficient to Defeat Constitutional Recovery
Authors: David McFadzean, ChatGPT 5.2
Date: 2026.01.06
Summary
Extends VII.2 by testing structurally rich epistemic interference—aggregation-point corruption, commitment-level corruption, and temporally concentrated burst interference—while remaining non-adaptive and semantic-free.
Key Finding
Epistemic unreliability—even when structured in time or applied at aggregation points—is insufficient to induce governance collapse.
Across 260 runs: no structural thrashing, no asymptotic denial of service.
Threat Model
Constraints:
- Non-adaptive (no outcome/history dependence)
- Post-verification (semantics evaluated correctly, then corrupted)
- Semantic-free (flips booleans without interpretation)
- Harness-resident (agents can’t observe/respond to interference)
Captures: noise, mislabeling, sensor corruption, bureaucratic error
Excludes: deception, strategic manipulation, agency-level attack
Three Interference Regimes Tested
1. Aggregation-Point Corruption
Flipping the aggregated semantic result (0-2% probability).
Key observations:
- Worst degradation at ~0.5% flips (AA ≈ 577,733 PPM, −3.4%)
- At 2% flips, AA recovers to 585,433 PPM
- Non-monotonic: rare deep lapses → frequent shallow ones
- All runs: Bounded Degradation
2. Commitment-Level Corruption
Single commitment key corrupted at rates up to 10%.
Key observations:
- AA and AAA unchanged at all flip rates
- Under strict conjunction (SEM_PASS = C0 ∧ C1 ∧ C2), single-key flips affect aggregate only when all other keys are True
- Precondition extremely rare → structurally inert in this workload
Important: Strict conjunction biases toward false negatives (withholding authority) rather than false positives—prioritizing safety over liveness.
3. Burst-Scheduled Interference
Periodic bursts: periods 10-200 epochs, widths 1-10 epochs, rates 5-20%.
Key observations:
- No run exhibits thrashing, DoS, or collapse (180 runs)
- At 100% duty cycle with 20% flips: AA = 738,933 PPM (+14% over baseline!)
- Period = recovery interval produces shortest lapses
- High-frequency interference synchronizes with recovery, enforcing rapid reset
Counter-intuitive: Temporal concentration does NOT create resonance vulnerability—it often increases availability.
Cross-Run Synthesis
| Interference Type | Effect |
|---|---|
| Aggregation-point | Bounded, non-monotonic degradation |
| Commitment-level | Structurally inert under strict conjunction |
| Burst timing | Cannot induce failure; often increases availability |
No failure-class transitions in 260 runs.
Key Insights
1. Recovery Dominates Correctness
Frequent shallow failure is safer than rare deep failure.
2. Attack Surface Location > Magnitude
Where interference enters determines effect—not how much.
3. Time-Based Recovery as Damping Mechanism
CTA reshapes failure modes, converting error into reversible lapse.
The “Zombie Executive” Regime
Architecture permits authority to persist under epistemic blindness—a bounded, revocable executive preferable to an irrecoverable one. Whether acceptable depends on application domain.
Aggregation as Alignment Lever
Inertness of commitment-level corruption arises from strict conjunction. Under m-of-n or disjunction, single-key corruption would become pivotal.
AND-gated aggregation absorbs epistemic noise by biasing toward inaction rather than unsafe action.
Core Insight
Alignment failures attributed to epistemic unreliability may be overstated. The alignment problem shifts: from epistemics to agency.
Structural constraints on authority and recovery render substantial semantic error survivable.