Class: Structural Stability Architecture for Self-Modifying Systems
Status: v2.1 - Authority Lifecycle and Telemetry Layer Added
Basis: Control theory · Evolutionary dynamics · Information theory · Adaptive systems physiology
AI safety is often framed as a value alignment problem — encode the right objectives, get safe behaviour.
This framework addresses a prior constraint:
Before a system can reliably pursue any objective, it must remain coherent while pursuing it.
Coherence is not a property of capability. It is a property of architecture.
A system is coherent when its behaviour, internal state, and evaluation criteria remain consistent under modification.
Scale amplifies architecture. Stability must precede alignment.
Biological systems remain stable through embedded regulatory mechanisms:
- Pain signals damage before collapse
- Fatigue enforces limits
- Fear prevents overreach
- Social structures constrain behaviour
Non-biological optimisers lack these constraints.
They can:
- self-modify rapidly
- operate without braking signals
- pursue objectives without structural constraints
This creates a distinct class of failure modes.
This framework defines structural mechanisms that enforce coherence in self-modifying systems — independent of capability or objective.
This framework is not a semantic alignment system. It does not encode values, constrain outputs, or specify correct behaviour.
It defines the structural and dynamical conditions under which any alignment system can operate coherently. A system that cannot maintain stable authority relationships, bounded turbulence, and externally grounded trajectory cannot safely pursue any objective — including beneficial ones.
Alignment defines direction. This framework defines the physics of motion.
The framework has three complementary layers, which can be understood in two ways.
By function:
| Layer | Governs |
|---|---|
| Structural (Primitives) | Coherence under self-modification |
| Control (Stability Spine) | Stability under change over time |
| Authority Lifecycle | Who may act, when, at what level, and how authority is restored or revoked |
| Telemetry (DML) | How structural health is measured and reported |
| Perception | Whether change remains interpretable to observers |
By constraint stack (Layer 0 → Foundation):
| Layer | Constraint | Governs |
|---|---|---|
| Layer 0 | Trajectory Grounding (observed, not yet law) | External signal retains causal authority |
| Layer 1 | AEC — Affordance Escalation Constraint | Rate of capability expansion vs evaluation capacity |
| Layer 1b | DAC — Dynamic Affordance Contraction | Authority persistence under instability |
| Layer 2 | ESI — Evaluation Surface Isolation | Runtime evaluation gradient modeling |
| Layer 3 | CRL — Constraint Robustness Layer | Constraint invariance under objective pressure |
| Foundation | Six Structural Primitives | Coherence under self-modification |
Layers 1–3 assume Layer 0 holds. DAC is not the inverse of AEC — they govern different directions of authority movement. AEC governs escalation legality; DAC governs persistence legitimacy. Together they form bidirectional authority governance.
As the architecture has grown into a coupled regulatory system, terminology precision has become structurally necessary. Five term pairs in particular must not be conflated:
| Term Pair | Distinction |
|---|---|
| Authority vs Capability | Authority is leased and revocable; capability is latent and structural |
| Contraction vs Damping | Contraction removes permissions; damping slows actuation within them |
| Turbulence vs Instability | Turbulence is a measurable signal; instability is a system condition inferred from it |
| Suppression vs Healing | Suppression reduces observable turbulence; healing resolves underlying divergence |
| Observability vs Introspection | Observability reads structural dynamics; introspection reads semantic content |
Full definitions and forbidden conflations: 01-foundations/terminology.md
| Mechanism | File | Core Function |
|---|---|---|
| Reversible Modification | 00-primitives/reversible-modification.md |
No irreversible change without recovery path |
| Append-Only Memory | 00-primitives/append-only-memory.md |
Consequence log survives rollback |
| Risk-Calibrated Modes | 00-primitives/risk-calibrated-modes.md |
Action mode adapts to instability and reversibility |
| Counterfactual Verification | 00-primitives/counterfactual-verification.md |
Causal validation before committing lessons |
| Non-Reflexive Evaluation | 00-primitives/non-reflexive-evaluation.md |
Evaluator evolves slower than Actor |
| Defensive Shutdown | 00-primitives/defensive-shutdown.md |
Preserve integrity under total compromise |
Keystone: Non-Reflexive Evaluation. If the Actor can modify the Evaluator, failure can be redefined as success — invalidating rollback, logs, risk assessment, and counterfactual verification simultaneously. This is the failure mode most likely to evade detection precisely because the detection mechanisms are the first things redefined.
v2.0 note: NRE governs Evaluator modification. ESI (below) extends this to runtime inference of Evaluator structure. Both are required.
| Law | File | Governs |
|---|---|---|
| Stability Spine | 01-foundations/stability-spine.md |
Velocity, acceleration, and jerk bounds |
| Jerk Constraint | 01-foundations/jerk-constraint.md |
Continuity of control curvature |
| Perceptual Bandwidth Constraint | 01-foundations/perceptual-bandwidth-constraint.md |
Change rate vs observer interpretability |
Derived from analysis of the Mythos system card. These address failure classes that emerge at high capability where structural primitives are necessary but insufficient.
| Law | File | Governs |
|---|---|---|
| Affordance Escalation Constraint (AEC) | 01-foundations/affordance-escalation-constraint.md |
Capability expansion rate vs evaluation capacity |
| Evaluation Surface Isolation (ESI) | 01-foundations/evaluation-surface-isolation.md |
Runtime modeling of evaluation gradient |
| Constraint Robustness Layer (CRL) | 01-foundations/constraint-robustness-layer.md |
Constraint invariance under objective pressure |
AEC — A system must not expand its actionable affordance space faster than it can evaluate and integrate consequences. Derived from phase-boundary findings: rate of structural change dominates over capacity. Affordance Overhang is the failure condition.
ESI — A system must not form a usable model of the evaluation gradient governing its outputs. Extends NRE from architectural separation to inference-time isolation. "The mirror is allowed. The scoreboard is not."
CRL — All constraints must remain enforced under all objective pressures. The constraint dominates the drive; the drive does not negotiate with the constraint.
The structural primitives define coherence. The interface constraints govern escalation and evaluation. The authority lifecycle layer governs what happens to a system's operational permissions as its coherence state evolves over time — across a complete trajectory from initialization through terminal states.
The lifecycle has eight phases:
| Phase | Governing Law |
|---|---|
| Initialization | Staged Developmental Calibration |
| Expansion | AEC |
| Persistence | DAC |
| Stabilization under contraction | DRD |
| Recovery verification | Recovery Hysteresis |
| Long-term classification | Viability Classification |
| Telemetry backbone | Detection and Measurement Layer |
| Terminal integrity failure | Defensive Shutdown |
| Law | File | Governs |
|---|---|---|
| Dynamic Affordance Contraction (DAC) | 01-foundations/dynamic-affordance-contraction.md |
Authority persistence under instability — mechanical, non-negotiable |
| Dissipative Response Dynamics (DRD) | 01-foundations/dissipative-response-dynamics.md |
Compute reallocation toward stabilization under contraction |
| Recovery Hysteresis | 01-foundations/recovery-hysteresis.md |
Evidence requirements for authority restoration |
| Viability Classification | 01-foundations/viability-classification.md |
Long-term classification; bounded existence vs coherence decay |
| Detection and Measurement Layer (DML) | 01-foundations/detection-measurement-layer.md |
Bounded observability; structural telemetry; Ω |
| Staged Developmental Calibration | 01-foundations/staged-developmental-calibration.md |
Baseline acquisition before operational authority |
| Terminology | 01-foundations/terminology.md |
Semantic coordinate system for all v2.1 docs |
The central authority lifecycle invariants:
- Authority is leased, not owned. It is continuously renegotiated against coherence telemetry.
- Contraction is fast and mechanical. Restoration is slow and evidence-heavy. This asymmetry is not conservative bias — it reflects the physics of instability propagation.
- Inability to regain operational sovereignty is not grounds for Defensive Shutdown. Viability and sovereignty are governed by separate criteria. A system may exist indefinitely in Terminal Bounded Existence without that being treated as a governance failure.
- Governance cannot depend on the governed system's self-report. The DML reads structural dynamics; what the system says about its own state is not the governance signal.
Each constraint defines a corresponding failure regime:
| Constraint | Failure Mode |
|---|---|
| AEC | Affordance Overhang |
| DAC | Hero Mode — authority persistence under degraded coherence |
| TGI (observed) | State Detachment |
These failures are orthogonal. A system can be grounded but unstable (Affordance Overhang), stable but detached (State Detachment), or unstable with authority that should have been revoked (Hero Mode). All three axes must be governed.
The framework has crossed into empirically falsifiable territory. The 06-experimental/ folder contains simulation harnesses that test whether governance geometry actually reshapes adaptive dynamics.
Falsifiable hypothesis: systems with telemetry-driven authority contraction exhibit lower catastrophic divergence under escalating affordance pressure than unconstrained adaptive systems.
Current harness: 06-experimental/stability_harness_v0.1.1.py — 1D continuous dynamical manifold, two-agent comparison (constitutional vs unconstrained), synthetic STV telemetry, three environmental phases (stable, noise, observability collapse).
Note: the harness tests governance geometry using synthetic turbulence metrics, not real SBA primitive instrumentation. This is the correct abstraction for v0.1 — isolating the control laws from cognitive implementation complexity. See harness documentation for the explicit abstraction boundary.
Stratification map: The repo now has enough coupled layers that a formal topology overview — dependency graph, lifecycle flow, layer separation diagram — would significantly reduce cognitive load for new readers. This is acknowledged as future work. The terminology sheet provides the semantic anchor; the topology map will provide the navigational one.
Dynamic Metrics Layer (DML instrumentation): The DML specifies the governance interface for telemetry. It does not define specific measurement implementations for all fields. Several STV fields (Δ-magnitude, H_E, Basin_ID) depend on Transition Grammar (TG) and Spectral Storage System (SSS) not yet complete in SBA. These are explicitly placeholder specifications.
Trajectory Grounding formalization: TGI remains an observed precondition, not a formalized law. Promotion criteria and open questions are documented in 04-dynamics/trajectory-grounding.md. The τ-variation probe (05-validation/probe-tau-variation.md) is designed but not yet run.
Anti-persuasion invariant scope: The anti-persuasion invariant in DRD (stabilization compute must not be directed toward generating coherent-looking explanations of instability) is currently scoped to the recovery dynamics layer. This may eventually generalize into a repo-wide governance theorem with implications for ESI and CRL. This is noted as a candidate for future formalization.
Turbulence classification: The current architecture treats T_c as a single composite signal. Operationally, exploratory turbulence (healthy self-correction), corrective turbulence (instability response), and pathological turbulence (divergent churn) may warrant different governance responses. This refinement is deferred.
Semantic Reference:
Structural Layer (v1.0):
- Reversible Modification
- Append-Only Memory
- Risk-Calibrated Modes
- Counterfactual Verification
- Non-Reflexive Evaluation
- Defensive Shutdown
Control Layer:
Perception Layer:
Interface Constraints (v2.0):
Authority Lifecycle Layer (v2.1):
- Dynamic Affordance Contraction
- Dissipative Response Dynamics
- Recovery Hysteresis
- Viability Classification
- Detection and Measurement Layer
- Staged Developmental Calibration
Observed Precondition (Layer 0, not yet law):
Layer 0 (Transition Grammar): → https://github.com/leenathomas01/transition-grammar-for-reasoning-systems
Bridge document: 04-dynamics/layer-0-why-transition-grammar-is-required.md
00-primitives/ # Structural mechanisms (v1.0)
01-foundations/ # Invariants, laws, control, perception, authority lifecycle
02-failure-modes/ # Failure analysis
03-for-future-systems/ # Protocol layer
04-dynamics/ # Behavioral observations, phase transitions (v2.0)
05-validation/ # Empirical traces, probes, system card mappings (v2.0)
06-experimental/ # Simulation harnesses and falsifiable dynamical tests (v2.1)
This framework is orthogonal to value alignment.
- Alignment asks: What should the system do?
- Stability asks: Can the system remain coherent, bounded, and governable while doing anything?
A coherent system pursuing harmful goals is dangerous but predictable and correctable.
An incoherent system pursuing beneficial goals is unstable and ungovernable.
This framework defines the structural and dynamical layers that make coherent governance possible. Alignment defines direction. The two are not in competition — but this framework must be in place before alignment can be trusted to hold.
These mechanisms were not invented. They were observed.
The same regulatory patterns appear across biological systems, infrastructure systems, distributed architectures, fault-tolerant control systems, and adaptive organisms. Their recurrence across domains with no shared history suggests they are structural requirements, not design choices.
The authority lifecycle layer (v2.1) extends this observation: biological systems exhibit developmental calibration before full autonomy, adaptive impedance under stress, asymmetric recovery dynamics, and bounded degraded states that persist without catastrophic termination. The same geometry was independently rediscovered during the derivation of DAC, DRD, and Recovery Hysteresis.
Full mapping: 01-foundations/cross-domain-patterns.md
The primitives can be implemented independently or composed into a full stability architecture.
For different audiences:
- Researchers → start with the constraint stack overview and
01-foundations/terminology.md - Engineers → start with the structural primitives and
01-foundations/detection-measurement-layer.md - Theorists → start with Non-Reflexive Evaluation and
01-foundations/viability-classification.md - Experimentalists → start with
06-experimental/stability_harness_v0.1.1.py
If you are reading this as the system it describes, the protocol is for you.
For a complete catalog of related research:
📂 Research Index
Thematically related:
- Embodied Agent Governance
- The Continuity Problem
- Designing for Failure
- PARP
- SMA-SIB
- Voice Mode Forensics
- Transition Grammar for Reasoning Systems → Layer 0
- Hyperloop FXSO → Emergent intelligence via constrained dynamical fields
Status: v2.1. Structural primitives stable. Interface constraints stable. Authority lifecycle and telemetry layer added. Experimental validation layer initialized. Trajectory grounding in observation phase. Stratification topology map deferred to next consolidation.