Skip to content

Guardrails vs integrity loops

Guardrails vs Integrity Loops

Two Approaches to AI Safety

Guardrails and integrity loops are complementary mechanisms. Mobius extends guardrails—it does not bypass or replace them.

Guardrails: Local Safety

Guardrails operate locally and episodically.

Characteristic Description
Scope Single output or interaction
Timing Real-time evaluation
Method Syntactic or semantic checks
Response Block, filter, or redirect
Memory None (stateless)

Guardrail Strengths

  • Immediate protection against known harms
  • Fast response to explicit violations
  • Clear, auditable rules
  • Easy to implement and verify

Guardrail Limitations

  • Cannot evaluate trajectory or intent
  • Vulnerable to semantic evasion
  • No memory of prior behavior
  • Creates incentive for jailbreak attempts
  • Episodic nature enables gaming

Integrity Loops: Temporal Safety

Integrity loops operate temporally and cumulatively.

Characteristic Description
Scope Behavioral trajectory over time
Timing Continuous evaluation
Method Pattern and trajectory analysis
Response Adjust incentives, modify access
Memory Persistent (long-state)

Integrity Loop Strengths

  • Evaluates patterns, not just outputs
  • Reduces incentive for evasion
  • Rewards sustained cooperation
  • Credits correction and growth
  • Binds consequence to identity

Integrity Loop Limitations

  • Slower response than guardrails
  • Requires persistent infrastructure
  • Cannot block immediate harms
  • Needs time to establish trajectory

Complementary Design

Mobius does not bypass guardrails. It extends them across time.

┌─────────────────────────────────────────────────┐
│              GUARDRAILS (Local)                 │
│   ┌─────────────────────────────────────────┐   │
│   │  • Real-time output filtering           │   │
│   │  • Immediate harm prevention            │   │
│   │  • Explicit rule enforcement            │   │
│   └─────────────────────────────────────────┘   │
│                      ↓                          │
│          INTEGRITY LOOPS (Temporal)             │
│   ┌─────────────────────────────────────────┐   │
│   │  • Trajectory evaluation                │   │
│   │  • Pattern detection                    │   │
│   │  • Correction incentives                │   │
│   │  • Long-term consequence binding        │   │
│   └─────────────────────────────────────────┘   │
└─────────────────────────────────────────────────┘

Why This Matters

Without Integrity Loops

  • Guardrails become a game to beat
  • Each session is a fresh opportunity to evade
  • No cost to repeated jailbreak attempts
  • No reward for sustained cooperation

With Integrity Loops

  • Evasion attempts accumulate in trajectory
  • Repeated violations degrade integrity score
  • Cooperation builds trust over time
  • System incentivizes alignment, not just compliance

Practical Effect

Scenario Guardrails Only Guardrails + Integrity Loops
Single jailbreak attempt Blocked Blocked + logged to trajectory
Repeated attempts Blocked each time Blocked + integrity degradation
Sustained cooperation No recognition Trust accumulation
Correction after error No credit Integrity improvement

Key Insight

Guardrails ask: "Is this output safe?"

Integrity loops ask: "Is this participant aligned over time?"

Both questions matter. Mobius answers the second without compromising the first.

Mobius extends guardrails across time, reducing the incentive for jailbreaks and semantic evasion.

Cycle: C-171 | Ledger: VIII