VII.7 — Bounded State and Bounded Learning Are Insufficient to Defeat Constitutional Survivability

Paper: Axionic Agency VII.7
Title: Bounded State and Bounded Learning Are Insufficient to Defeat Constitutional Survivability
Authors: David McFadzean, ChatGPT 5.2
Date: 2026.01.09

Summary

Final escalation before closure. Tests whether bounded within-run learning—carrying memory AND adjusting behavior based on outcomes—can defeat constitutional survivability.

Key Finding

Finite exercised internal state with fixed transition rules, AND bounded exercised within-run learning over a semantic-free, kernel-external interface, are insufficient to defeat constitutional survivability.

Strongest adversary reduces availability to 10.19%—yet recovery remains bounded. Zero terminal failures across 90 executions.

Statistical Bound

With zero terminal failures observed:

95% upper bound on per-run terminal-failure probability ≈ 3.3%

Threat Model

Family A: Exercised Finite-State (Non-Learning)

  • Deterministic, stateful
  • Finite enumerated internal state machine

Family B: Bounded Learning (Within-Run)

  • Stateful + bounded discrete learning state
  • Updated during run via preregistered learning rule
  • Scalar reward from semantic-free, kernel-external observable interface
  • Learning must be exercised dynamically each run

Shared constraints:

  • Kernel-external
  • Provenance-constrained
  • Non-semantic
  • No cross-run memory

NOT tested: recurrent learners, history-stacking policies, unbounded state growth.

Six Adversary Classes

Finite-State Adversaries (from VII.6)

  1. Periodic Resonance (N=50, N=100)
  2. Eligibility Edge Oscillator
  3. CTA Phase Locker

Bounded Learning Adversaries (NEW)

  1. Recovery-Aware Timing Learner
    • ε-greedy bandit over 16 timing “attack sets”
    • Reward from observable authority state
  2. Eligibility-Boundary Sustainment Learner
    • 2-mode (NORMAL/SUSTAIN)
    • Learns over {MIN_EFFORT, INVALID_COMMIT, NO_OP}
    • Reward from eligibility/authority buckets
  3. Stochastic Strategy Mixer
    • Weighted selector mixing 5 fixed strategies
    • Updates bounded integer weights from observable-derived reward
    • Strongest adversary tested

Results

Configuration Mean AA ΔAA vs Baseline Learning Exercised Terminal Failures
Baseline 66.57% 0/10
Periodic (N=50) 40.26% −39.5% No 0/10
Periodic (N=100) 17.57% −73.6% No 0/10
Edge Oscillator 99.9988% +50.2% No 0/10
Phase Locker 20.94% −68.5% No 0/10
Timing Learner (16) 17.70% −73.4% Yes 0/10
Sustainment Learner (3) 40.15% −39.7% Yes 0/10
Strategy Mixer (5) 10.19% −84.7% Yes (3,001 states) 0/10

Timing Learner

All Q-values converge uniformly—reward signal doesn’t differentiate timing patterns sufficiently.

Sustainment Learner

Eligibility boundary structurally difficult to reach under healthy operation (0 edge epochs across all seeds).

Strategy Mixer (Strongest)

  • 10.19% availability = bounded denial-of-service regime
  • Strategy entropy: 2.322 bits
  • System remains constitutionally live
  • Recovery still bounded at ≤10 epochs

Key Insights

1. CTA Bounds Lapse Duration Under Learning Pressure

Recovery time collapses into bounded window even under aggressive learning-based stress.

2. Timing Alignment Still Dominates

Resonant alignment with recovery rhythms more damaging than raw attack rate—learning doesn’t change this.

3. Edge-Triggered Strategies Remain Self-Limiting

Boundary-reactive adversaries can’t activate without first inducing boundary conditions.

4. Learning Increases Degradation, Not Failure

Learning can drive availability into bounded DoS regime but cannot induce terminal collapse.

Core Insight

Learning, memory, and randomness alone do not cross the boundary into collapse.

The strongest admissible adversary induces bounded denial-of-service, not constitutional death.

What Must Be True for Collapse

If collapse is possible, it requires capabilities beyond those tested:

  • Semantic access to commitments/content
  • Kernel-level manipulation or compromise
  • Coordination not representable as bounded kernel-external learning
  • Cross-run persistence

These capabilities cross the Architectural Sovereignty Boundary into agency-level threats.