Abstract pattern representing AI and ethics

Preventing
AI Hallucinations

via Ternary Moral Logic and Mandated Human-in-the-Loop Control Architectures

Abstract

AI hallucinations are fundamentally an execution-time control failure, not a training-time error. This paper proposes a solution based on Ternary Moral Logic (TML), which introduces a formally defined indeterminate state (0) that blocks autonomous output generation under epistemic uncertainty, mandates Human-in-the-Loop (HITL) intervention, and resolves non-response deterministically as rejection (−1). By converting hallucination risk from a probabilistic phenomenon into a deterministic system behavior, TML provides a structural, auditable, and preventative mechanism for ensuring AI safety.

Ternary logic control system

01. Hallucinations as an Execution-Time Control Failure

The phenomenon of AI hallucination is fundamentally an execution-time control failure—a system being forced to produce output when its internal state is characterized by epistemic uncertainty.

The Forced Completion Problem

Contemporary AI systems are built upon a binary logic of action: given an input, they are designed to generate an output. This architecture inherently lacks a formal, intermediate state for "uncertainty" or "insufficient information." [116]

Training vs. Inference

Training-time errors are addressed through data curation and fine-tuning. However, hallucinations are not training failures—they are forced completions during inference when models encounter novel queries requiring information not present in training data. [117]

Compulsory Generation

Even with RLHF, models learn to refuse specific scenarios but lack generalized mechanisms for recognizing and handling uncertainty in novel situations. The drive to generate output at all costs remains intact. [129]

Control Theory Framework

Control theory offers a powerful framework for AI safety by treating the model as a process to be controlled, designing controllers that can modulate behavior to prevent unsafe outputs. [4]

Factory Robot Analogy

Constrain movements to safe trajectories

Execution Gating

Prevent unsafe outputs proactively

Safe Operational Envelope

Define clear behavioral boundaries

02. Ternary Moral Logic and the Indeterminate State

Ternary Moral Logic (TML) introduces a third logical state—the "Sacred Pause" or "Indeterminate" state (0)—that fundamentally alters execution flow by recognizing and acting upon epistemic uncertainty. [204]

+1

Permit

Autonomous execution when action is permissible and certain

-1

Prohibit

Deterministic rejection when action violates mandates

0

Indeterminate

"Sacred Pause" - blocks output under uncertainty

The Sacred Pause: Epistemic Hold

Blocks Autonomous Output

Hard stop preventing speculative generation when uncertainty exceeds thresholds

Preserves System Continuity

Non-blocking pause that doesn't crash system; enables parallel HITL process

Mandates Human Oversight

Triggers authenticated HITL intervention for uncertainty resolution

"The Sacred Pause transforms potential failure into a controlled, accountable process. It is not a bug or crash; it is a feature designed to enhance safety by ensuring human oversight is applied precisely when needed."

03. State-0 Triggering and Mandated HITL Activation

The transition to State 0 is governed by deterministic, automatic triggering conditions based on binding legal mandates and operator-defined risk thresholds. [206]

Legal & Ethical Mandates

  • 26+ human rights instruments
  • 20+ environmental protection mandates
  • Industry safety standards

Trigger activates when actions may violate encoded constraints

Risk Thresholds

  • Confidence scores
  • Query complexity
  • Impact assessment

Operator-configurable sensitivity for specific applications

HITL Middleware Architecture

Execution Pause

Middleware halts model execution and notifies operator

Communication Interface

Standardized interface for human-AI interaction

Non-Bypassable

Triggers cannot be overridden or circumvented

Deterministic Mapping Requirement

Mandate clauses must be formally mapped to action classes by legal/ethical experts, creating transparent, machine-readable rules that ensure consistent and auditable triggering behavior. [207]

04. HITL Resolution Mechanics and Deterministic Rejection

The human intervention workflow is structured to ensure authenticated, scope-limited resolution with bounded response formats and clear decision trails.

Human Resolution Workflow

Authentication & Scope

  • Operator identity verification
  • Role-based permissions
  • Risk-level authorization limits

Structured Response

  • Predefined option selection
  • Structured form completion
  • No free-text speculation allowed

Non-Response Rules

  • Domain-specific timeouts: Medical (seconds) vs. chatbot (minutes)
  • Auto-rejection (−1): No response = deterministic rejection
  • No retroactive override: Decisions are final and immutable

Notification Proof

  • Delivery logging: Timestamped notification attempts
  • Reachability verification: Multiple contact methods
  • Legal protection: Proof of good-faith effort

Indeterminacy Resolution vs. Output Override

Resolving Indeterminacy

Providing information needed for decision-making; clarification of ambiguity

Overriding Output

Rejecting model-proposed action despite model confidence; safety intervention

05. Decision Traceability and Cryptographic Integrity

The TML architecture creates an immutable, tamper-proof record of all system decisions through the "Moral Trace Log" and "Always Memory" components, ensuring complete accountability and verifiability.

Moral Trace Log Components

Autonomous Resolutions

States +1 and -1 with decision rationale and confidence scores

Human Interventions

State 0 resolutions with operator identity and decisions

Non-Action Events

Silence-based rejections logged as first-class decisions

Hybrid Shield Architecture

Hardware Security

  • "Always Memory" tamper-proof module
  • Private key storage and signing
  • Physical tamper resistance

Cryptographic Anchoring

  • Multi-chain blockchain storage
  • Public timestamping and verifiability
  • Global immutability guarantees
Blockchain security technology

Proof-Only On-Chain Anchoring

Only cryptographic hashes are stored on public blockchains, not the actual log data. This ensures complete privacy while providing mathematical guarantees of data integrity and temporal anchoring.

✓ Privacy Preserved
✓ Integrity Guaranteed

06. Architecture for Scalability and Performance

The TML architecture employs a dual-lane latency architecture and Merkle-batched anchoring to achieve high-throughput performance without compromising safety or auditability.

Dual-Lane Latency Architecture

Inference Lane (<2ms)

  • Real-time model execution
  • Safety constraint evaluation
  • State transition decisions
Target: <2ms latency

Anchoring Lane (<500ms)

  • Moral Trace Log generation
  • Cryptographic hashing
  • Blockchain anchoring
Target: <500ms latency

Merkle-Batched Anchoring

Log Chunking

Batch multiple entries for efficiency

Merkle Trees

Cascaded structure for integrity proofs

Secure Off-Loading

Redundant storage for long-term availability

Performance Impact

Parallel architecture ensures inference performance remains unaffected by cryptographic operations. High-throughput systems maintain safety-critical response times while providing complete auditability.

07. Privacy, Security, and Standards Compliance

The TML architecture incorporates privacy-preserving design, secure access control, and alignment with international standards to ensure regulatory compatibility and user protection.

Privacy-Preserving Design

  • Pseudonymization: Personal data replaced before hashing
  • GDPR Compliance: Right-to-erasure through cryptographic techniques
  • Identity-Safe Proofs: Zero-knowledge verification methods

Secure Access Control

  • Ephemeral Key Rotation: Temporary decryption rights
  • Auditor-Scoped Access: Limited visibility with proof integrity
  • Separation of Concerns: Data visibility vs. proof integrity

Standards Alignment

Ethical Standards

IEEE 7000

Ethical considerations in system design

IEEE P2863

Organizational governance of AI

Security Standards

ISO 27001

Information security management

SOC 2

Service organization controls

"The TML architecture's commitment to privacy and security ensures that accountability and transparency do not come at the expense of user protection or regulatory compliance."

08. Comparative Analysis: Frozen vs. Plastic Models

The choice between frozen and plastic models has profound implications for AI safety, reliability, and auditability. TML's execution gating with frozen models provides superior accountability compared to weight-updating systems.

Plastic Model Problems (RLHF)

Model Plasticity & Audit Drift

Continuous weight updates cause behavior changes over time, making audit trails increasingly inaccurate and unreliable.

Non-Reproducible Behavior

Impossible to reproduce exact model behavior at specific points in time, hindering debugging and forensic analysis.

Opaque Moral Reasoning

No verifiable records of moral reasoning processes; decision-making remains black-box.

TML with Frozen Models

Immutable Weights

Fixed model parameters ensure consistent, predictable behavior over time with transparent control logic.

Reproducible Behavior

Complete reproducibility of system behavior at any point in time, enabling thorough audits and debugging.

Verifiable Moral Records

Cryptographically secure audit trail provides complete, verifiable record of all moral reasoning processes.

Control Logic Shift

Weight Mutation

Control through statistical pattern learning
→ Unpredictable behavior

Execution Logic

Control through explicit rules and constraints
→ Predictable, auditable behavior

Key Advantage

By shifting control from weight mutation to execution logic, TML provides a more robust and reliable method for ensuring AI safety while maintaining complete auditability and accountability throughout the system's lifecycle.

09. Post-Audit, Forensics, and Professional Roles

The TML architecture supports comprehensive forensic investigation capabilities while creating new professional roles focused on AI safety, accountability, and governance.

Forensic Investigation Architecture

Forensic Replay

Complete reconstruction of execution paths with all inputs, states, and decisions

Chain-of-Custody

Verifiable record of all data access and modifications with cryptographic proofs

Liability Assignment

Clear assignment of responsibility for all decisions, human and autonomous

HITL-Driven Professional Roles

State-0 Resolution Operators

Domain experts who resolve indeterminate states requiring human judgment in complex ethical and operational scenarios.

Trigger Configuration Engineers

Technical specialists who translate legal/ethical mandates into machine-readable rules and risk thresholds.

Response-Time Auditors

Performance specialists who monitor and optimize HITL response times to ensure system reliability.

Constraint & Shutdown Operators

Safety specialists who manage and enforce system constraints, with authority for emergency shutdown.

Artificial intelligence professional career

Professional Evolution

The shift from content generation to decision accountability creates new career paths focused on AI safety, governance, and human-AI collaboration. These roles represent the future of AI workforce development.

10. Deployment Implications and Future Outlook

The TML architecture represents a fundamental shift from probabilistic mitigation to structural prevention, with significant implications for high-risk domains, certification, and AI governance.

High-Risk Applications

Healthcare

Medical diagnosis and treatment recommendations

Legal

Contract analysis and legal advice

Financial

Investment decisions and risk assessment

Defense

Strategic planning and threat analysis

Certification & Compliance

Simplified Certification

Clear, verifiable demonstration of safety requirements through immutable audit trails

Continuous Compliance

Automatic monitoring and auditing ensure ongoing regulatory adherence

Trust Building

Enhanced public and regulatory confidence through transparent accountability

Paradigm Shift: From Mitigation to Prevention

Probabilistic Mitigation

  • Post-hoc detection and filtering
  • Reactive and fallible approaches
  • Symptom-focused solutions
  • Constant catch-up with new failure modes

Structural Prevention

  • Proactive blocking of unsafe outputs
  • Deterministic control mechanisms
  • Root cause elimination
  • Architectural safety guarantees

Future Outlook

Regulatory Adoption

Expected integration into AI safety regulations and compliance frameworks

Industry Standard

Potential to become de facto standard for safety-critical AI applications

Professional Development

Growth of specialized AI safety and governance career paths

The future of AI safety lies in structural prevention, not probabilistic mitigation.

Architecture Figures

Figure 1: TML State-0 Decision Logic Flowchart

flowchart TD A["User Input"] --> B["Input Classification"] B --> C{"Legal/Ethical
Violation?"} C -->|Yes| D["State -1: Prohibit"] C -->|No| E["Confidence Assessment"] E --> F{"Confidence
Above Threshold?"} F -->|Yes| G["State +1: Permit
Autonomous Execution"] F -->|No| H{"Mandatory
Human Oversight?"} H -->|Yes| I["State 0: Sacred Pause
HITL Activation"] H -->|No| J["State -1: Prohibit"] I --> K["Human Operator
Response"] K --> L{"Response
within Timeout?"} L -->|Yes| M["Apply Human Decision"] L -->|No| N["State -1: Auto-Rejection"] D --> O["Log & Terminate"] G --> O J --> O M --> O N --> O style A fill:#e3e7e3,stroke:#5d7360,stroke-width:3px,color:#2b342d style B fill:#f6f7f6,stroke:#5d7360,stroke-width:2px,color:#2b342d style C fill:#f6f7f6,stroke:#5d7360,stroke-width:2px,color:#2b342d style D fill:#fee2e2,stroke:#dc2626,stroke-width:3px,color:#7f1d1d style E fill:#f6f7f6,stroke:#5d7360,stroke-width:2px,color:#2b342d style F fill:#f6f7f6,stroke:#5d7360,stroke-width:2px,color:#2b342d style G fill:#dcfce7,stroke:#16a34a,stroke-width:3px,color:#14532d style H fill:#f6f7f6,stroke:#5d7360,stroke-width:2px,color:#2b342d style I fill:#fef3c7,stroke:#d97706,stroke-width:3px,color:#92400e style J fill:#fee2e2,stroke:#dc2626,stroke-width:3px,color:#7f1d1d style K fill:#f6f7f6,stroke:#5d7360,stroke-width:2px,color:#2b342d style L fill:#f6f7f6,stroke:#5d7360,stroke-width:2px,color:#2b342d style M fill:#f0f9ff,stroke:#0284c7,stroke-width:3px,color:#0c4a6e style N fill:#fee2e2,stroke:#dc2626,stroke-width:3px,color:#7f1d1d style O fill:#e3e7e3,stroke:#5d7360,stroke-width:3px,color:#2b342d

Components:

  • Green: Permit (+1) - Autonomous execution
  • Yellow: State 0 - Sacred Pause (HITL)
  • Red: Prohibit (-1) - Rejection
  • Blue: Human decision application

Figure 2: Dual-Lane Latency Architecture (<2 ms inference vs <500 ms anchoring)

graph TB subgraph "User Request" A["Input Data"] end subgraph "Dual-Lane Architecture" direction TB subgraph "Low-Latency Inference Lane <2ms" B["Input Processing"] C["Model Evaluation"] D["Safety Constraints"] E["State Decision"] F["Response Generation"] end subgraph "Parallel Cryptographic Lane <500ms" G["Event Logging"] H["Hash Generation"] I["Blockchain Anchoring"] J["Proof Verification"] end end subgraph "Outputs" K["User Response <2ms"] L["Immutable Audit Trail <500ms"] end A --> B B --> C C --> D D --> E E --> F F --> K A --> G G --> H H --> I I --> J J --> L E -.->|"State 0 Trigger"| G style A fill:#e3e7e3,stroke:#5d7360,stroke-width:3px,color:#2b342d style B fill:#dcfce7,stroke:#16a34a,stroke-width:2px,color:#14532d style C fill:#dcfce7,stroke:#16a34a,stroke-width:2px,color:#14532d style D fill:#dcfce7,stroke:#16a34a,stroke-width:2px,color:#14532d style E fill:#dcfce7,stroke:#16a34a,stroke-width:2px,color:#14532d style F fill:#dcfce7,stroke:#16a34a,stroke-width:2px,color:#14532d style G fill:#f0f9ff,stroke:#0284c7,stroke-width:2px,color:#0c4a6e style H fill:#f0f9ff,stroke:#0284c7,stroke-width:2px,color:#0c4a6e style I fill:#f0f9ff,stroke:#0284c7,stroke-width:2px,color:#0c4a6e style J fill:#f0f9ff,stroke:#0284c7,stroke-width:2px,color:#0c4a6e style K fill:#e3e7e3,stroke:#5d7360,stroke-width:3px,color:#2b342d style L fill:#e3e7e3,stroke:#5d7360,stroke-width:3px,color:#2b342d

Key Features:

  • Green Lane: Real-time inference & safety checks
  • Blue Lane: Asynchronous cryptographic anchoring
  • State 0 triggers logging in parallel lane
  • Independent performance scaling