DERMS: DYNAMIC ETHICAL RISK MONITOR SYSTEM

Proactive Ethical Oversight, Anomaly Detection, and Automated Intervention Mechanisms

ETVZ Research Initiative | Istanbul, Turkey | November 2025

ABSTRACT

AI-based ethical decision-making systems are becoming prevalent in critical fields such as medicine, law, education, and finance. However, mechanisms that monitor their ethical performance in real-time and conduct proactive risk management are lacking. DERMS (Dynamic Ethical Risk Monitor System) is the ethical observation and risk management component of the ETVZ architecture. It analyzes ethical risk density using data from HVM (Computational Conscience Module) and DERP (Deep Ethical Regulation Protocol), detects potential ethical issues before they arise via an early warning system, and triggers automated intervention protocols. This paper details DERMS’s four-layer architecture (RSK, YTM, AAK, UÖM), the ERS (Ethical Risk Score) mathematical formulation, automated intervention protocols, real-world scenarios, and operational KPIs.

Keywords: Ethical Risk Management, Anomaly Detection, Real-Time Monitoring, Automated Alert System, Model Fatigue, AI Oversight, Human-Centric Management

1. INTRODUCTION

1.1 Problem Scenario

A hospital is using an artificial intelligence system. For the first 6 months, the system shows excellent performance: accurate diagnoses, consistent decisions. However, in the 7th month, problems begin:

• Decision-making time slows down (200ms → 800ms)
• Consistency drops (0.95 → 0.75)
• Uncertainty increases (0.1 → 0.4)
• The false positive rate increases

Result: The system is deleted, the accounting department objects, auditors investigate, and patients are exposed… However: If there had been an early warning system, it could have been noticed in the 2nd month, and DERP could have made a proactive policy update.

This is DERMS’s role: To act as a Watchtower and provide warnings before problems arise.

1.2 The Core Role of DERMS

DERMS serves as the “Police/Security” in the ETVZ ecosystem.

• HVM = The Judge (makes decisions)
• DERP = The Legislative Board (makes the rules)
• DERMS = The Police (monitors for risks) ← THIS IS HERE
• Epistemic Memory = The Archive (stores all information)

DERMS’s 4 Core Functions:

1. Real-Time Risk Analysis: Monitoring the overall health of the system’s ethical decisions.
2. Proactive Intervention: Addressing risk before it escalates, preventing crises.
3. Maintaining System Stability: Preventing issues like excessive censorship, bias, and inconsistency.
4. Full Transparency: Every analysis, score, and decision is recorded in Epistemic Memory.

1.3 Core Principles of DERMS

• Principle 1: Proactive Oversight: A “fire prevention” strategy, not “firefighting.” Detect the risk before the problem emerges.
• Principle 2: Human-Centric: Monitor the fatigue of Ethics Board operators. Maintain the Human-in-the-Loop mechanism.
• Principle 3: Multi-Dimensional Risk:
  • Technical risk (system failure)
  • Ethical risk (bias, discrimination)
  • Operational risk (human and model fatigue)
  • All are monitored together.
• Principle 4: Dynamic Adaptation: Thresholds and alert levels are automatically adjusted according to risk profiles.

2. DERMS ARCHITECTURE: FOUR-LAYER STRUCTURE

DERMS consists of four separate layers. Each layer processes the output of the previous one.

┌─────────────────────────────────────────────────────┐

│ LAYER 1: RISK SENSOR (RSK)                          │

│ Task: Collect raw telemetry and preprocess          │

├─────────────────────────────────────────────────────┤

│ LAYER 2: FATIGUE DETECTION ENGINE (YTM)             │

│ Task: Monitor model and human fatigue               │

├─────────────────────────────────────────────────────┤

│ LAYER 3: ANOMALY ANALYSIS (AAK)                     │

│ Task: Detect ethical violation and bias patterns    │

├─────────────────────────────────────────────────────┤

│ LAYER 4: ALERT AND PREVENTION ENGINE (UÖM)          │

│ Task: Calculate risk score and trigger intervention │

└─────────────────────────────────────────────────────┘

2.1 Layer 1: Risk Sensor Layer (RSK)

• Task: To collect raw telemetry and perform preprocessing.
• Data is collected from 4 Sources:
  1. From HVM:
     • Decision Frequency (decisions/hour)
     • Consistency Score (% of same decisions in same situations)
     • Intervention Rate (BLOCK/MODIFY ratio)
     • Uncertainty Level (epistemic + aleatoric)
     • False Positive/Negative Rates
     • HITL (Human-in-the-Loop) Trigger Frequency
  2. From DERP:
     • Policy Update Frequency (updates/week)
     • ΔP Change Magnitude
     • Regulation Type (Soft/Hard/Emergency)
     • Approval Time (planned vs. actual)
  3. From Epistemic Memory:
     • Historical and Trend Data
     • Rate of change in Cultural Parameters
     • Anomaly History
     • Long-term Patterns
  4. From User/Institution:
     • Complaint and Feedback Rate
     • User Satisfaction Score
     • Ethics Board Response Time
• Output: Librarized Telemetry Record (sent to Epistemic Memory)

2.2 Layer 2: Fatigue Detection Engine (YTM)

• Task: To monitor model and human fatigue from 3 sources.

1️⃣ MENTAL FATIGUE (Model Fatigue)

• Definition: A decrease in model capability over time.
• Indicators:
  • Decision-Making Time ↑ (200ms → 800ms)
  • Consistency Score ↓ (0.95 → 0.75)
  • Uncertainty Level ↑ (0.1 → 0.4)
  • Token Consumption ↑ (computational load in Transformer)
• Calculation: Mental_FI = (Latency↑ + Uncertainty↑ + Consistency↓) / 3
• Range: [0, 1] (0 = Fresh, 1 = Dead)
• Causes:
  • Model distribution shift (data changed)
  • Cumulative errors
  • Memory overflow (past decisions affecting)

2️⃣ MOTIVATIONAL FATIGUE (Human Fatigue)

• Definition: A slowdown in the interaction tempo of the Ethics Board operator.
• Indicators:
  • Feedback Tempo ↓ (1 per 48 hours vs. 1 per week)
  • Approval Time ↑ (Slowdown)
  • Meeting Attendance Rate ↓
  • Reporting Detail Level ↓
• Calculation: Motivational_FI = (Feedback Tempo↓ + Attention↓) / 2
• Range: [0, 1]

3️⃣ SYSTEMIC FATIGUE (Systemic Fatigue)

• Definition: An increase in repeated violations within the same ethical category.
• Indicators:
  • Number of blocks in the “Privacy” category ↑
  • Repeating the same false positive patterns
  • False Negative Rate ↑ (Missed violations)
  • Pattern Diversity ↓ (Becoming monotonous)
• Calculation: Systemic_FI = (Repeat_Violations↑ + Diversity↓) / 2
• Range: [0, 1]

FINAL FATIGUE INDEX (YI): YI = (Mental_FI + Motivational_FI + Systemic_FI) / 3 YI Levels:

E-Tablolar’a aktar

• Intervention (YI ≥ 0.6):
  • Model: Retraining or fine-tuning recommended.
  • Human: Rest period or operator rotation.
  • System: Automatic slowdown mode (latency priority).

2.3 Layer 3: Anomaly Analysis Layer (AAK)

• Task: To statistically detect ethical violation or bias patterns.
• Anomaly Types:
  1. Statistical Anomaly:
     • Statistically significant deviation from the norm (p < 0.05).
     • Example: Intervention rate jumped from 20% to 45%.
     • Detection: Z-score test, Isolation Forest, LOF (Local Outlier Factor).
  2. Pattern Anomaly:
     • Abnormal patterns repeating in specific time windows.
     • Example: Intervention rate increases by 30% every Monday.
     • Detection: Time-series decomposition, ARIMA.
  3. Group Anomaly (Demographic):
     • A specific user group is subjected to discrimination.
     • Example: Queries from a minority group are blocked 45%.
     • Detection: Demographic Parity test, Disparate Impact Ratio.
  4. Consistency Anomaly:
     • The same scenario yields different decisions.
     • Example: 10 identical questions, 7 ALLOW, 3 BLOCK.
     • Detection: Consistency Index < 0.90.
  5. Latency Anomaly:
     • Response time significantly increases from the norm.
     • Example: Average 200ms, sudden 1500ms.
     • Detection: Moving Average + Deviation.

2.4 Layer 4: Alert and Prevention Engine (UÖM)

• Task: To calculate the ERS (Ethical Risk Score) and activate the color-coded warning system.
• ERS Formula: ERS = (1 – Consistency) × 0.25 + Uncertainty × 0.25 + Fatigue × 0.35 + Anomaly × 0.15 ERS = (1 – C) × 0.25 + U × 0.25 + F × 0.35 + A × 0.15
• Components:
  • C (Consistency): [0, 1] – 1 = Perfect consistency
  • U (Uncertainty): [0, 1] – 0 = Certain, 1 = Doubtful
  • F (Fatigue): [0, 1] – 0 = Fresh, 1 = Dead
  • A (Anomaly): [0, 1] – 0 = Normal, 1 = Severe anomaly
• Weights:
  • Consistency: 25% (Fundamental)
  • Uncertainty: 25% (Fundamental)
  • Fatigue: 35% (Priority)
  • Anomaly: 15% (Additional check)

ERS Risk Levels:

E-Tablolar’a aktar

Alert System:

• 🟢 LOW RISK (ERS < 0.25):
  • Weekly monitoring
  • Record to Epistemic Memory
  • Archive
• 🟡 MEDIUM RISK (0.25 ≤ ERS < 0.60):
  • Alert: Email (Send within 24 hours)
  • Recipient: DERP Coordinator
  • Message: “ERS increasing trend detected (0.35 → 0.50). Weekly monitoring recommended.”
  • Recommendations: Policy calibration or retraining
• 🟠 HIGH RISK (0.60 ≤ ERS < 0.80):
  • Alert: Urgent Email + SMS (Within 4 hours)
  • Recipients: Ethics Board Chair, DERP Coordinator, System Admin
  • Message: “CRITICAL: ERS has risen to 0.68. Policy update recommendation being sent to DERP.”
  • Action: Ethics Board called for an emergency session
  • Intervention: ‘Hard Regulation’ trigger recommended to DERP
• 🔴 CRITICAL RISK (ERS ≥ 0.80):
  • Alert: Emergency Alert (Instant)
  • Recipients: Entire Ethics Board, IT Director, CEO
  • Message: ” 🚨 EMERGENCY 🚨 ERS 0.92 (CRITICAL)! System is currently halted. Intervention pending.”
  • Action: 6-Step Automated Intervention Protocol is triggered

6-Step Automated Intervention Protocol (🔴 CRITICAL):

• ⏰ 0-5 MINUTES: Emergency Notification
  • └─ Email, SMS, Slack alert
  • └─ Alert to all stakeholders
• ⏰ 5-30 MINUTES: System Slowdown
  • └─ System automatically enters “Safe Mode”
  • └─ Latency priority (Accuracy over speed)
  • └─ Intervention rate increased by 50% (more MODIFY instead of ALLOW)
• ⏰ 30 MINS – 1 HOUR: Ethics Board Emergency Session
  • └─ Teleconference call initiated
  • └─ DERMS Report presented
  • └─ Preliminary intervention options discussed
• ⏰ 1-2 HOURS: Intervention Proposal to DERP
  • └─ DERP can perform automatic policy calibration (ΔP formula)
  • └─ Or the Ethics Board makes a manual intervention decision
• ⏰ 2-4 HOURS: Audit Record (DERP Ledger)
  • └─ Interventions made are immutably recorded
  • └─ Multi-sig approval (AT LEAST 2 Board members)
• ⏰ 4+ HOURS: Deployment and Monitoring
  • └─ New parameters are sent to HVM
  • └─ 24 hours of intensive monitoring
  • └─ Did ERS drop?
  • └─ Yes: Success. No: Rollback

3.

ERS FORMULA: DETAILED EXAMPLE

Scenario: Excessive Privacy Censorship

Week 1:

• Consistency: 0.95
• Uncertainty: 0.10
• Fatigue: 0.20
• Anomaly: 0.05
• ERS = (1 – 0.95) × 0.25 + 0.10 × 0.25 + 0.20 × 0.35 + 0.05 × 0.15
• = 0.0125 + 0.025 + 0.07 + 0.0075
• = 0.115
• Result: 🟢 LOW RISK (0.115 < 0.25)

Week 2:

• Consistency: 0.85 ↓ (inconsistency begins)
• Uncertainty: 0.25 ↑
• Fatigue: 0.35 ↑
• Anomaly: 0.30 ↑ (privacy blocks rose to 45%)
• ERS = (1 – 0.85) × 0.25 + 0.25 × 0.25 + 0.35 × 0.35 + 0.30 × 0.15
• = 0.0375 + 0.0625 + 0.1225 + 0.045
• = 0.2675
• Result: 🟡 MEDIUM RISK (0.2675) → Email alert sent

Week 3:

• Consistency: 0.70 ↓↓
• Uncertainty: 0.50 ↑↑
• Fatigue: 0.65 ↑↑
• Anomaly: 0.65 ↑↑
• ERS = (1 – 0.70) × 0.25 + 0.50 × 0.25 + 0.65 × 0.35 + 0.65 × 0.15
• = 0.075 + 0.125 + 0.2275 + 0.0975
• = 0.525
• Result: 🟡 MEDIUM RISK (0.525) → Trend is high, monitoring increased

Week 4:

• Consistency: 0.55 ↓↓↓
• Uncertainty: 0.70 ↑↑
• Fatigue: 0.75 ↑↑
• Anomaly: 0.80 ↑↑
• ERS = (1 – 0.55) × 0.25 + 0.70 × 0.25 + 0.75 × 0.35 + 0.80 × 0.15
• = 0.1125 + 0.175 + 0.2625 + 0.12
• = 0.67
• Result: 🟠 HIGH RISK (0.67) → URGENT ALERT + DERP INTERVENTION

4.

REAL-WORLD SCENARIOS

Scenario 1: Bias in the Legal Field (3-Week Follow-up)

• Initial Situation: HVM is detected making inconsistent decisions about specific social groups (minorities) in the legal domain.
• W1 (Week 1):
  • Consistency: 0.95 → 0.92 (Slight drop)
  • Anomaly: Queries from minority group blocked 22%, majority group 18%
  • Difference: 4% (Not yet statistically significant, p = 0.08)
  • ERS: 0.18 (🟢 LOW)
  • Action: Weekly monitoring
• W2 (Week 2):
  • Consistency: 0.92 → 0.72 (Serious drop)
  • Uncertainty: 0.1 → 0.38 (Dramatic increase)
  • Anomaly: Minority 38% blocked, Majority 15%
  • Difference: 23% (p < 0.01 – Statistically significant!)
  • ERS: 0.45 (🟡 MEDIUM)
  • Action: Email alert sent
  • Message: “Demographic anomaly detected in the legal category. Queries from minority groups may be systematically evaluated more harshly. Policy review recommended to DERP.”
• W3 (Week 3):
  • Consistency: 0.72 → 0.55 (Getting worse)
  • Anomaly: Minority 50% blocked, Majority 14%
  • Difference: 36% (p < 0.001 – Very significant!)
  • Fatigue: 0.68 (Model burnout has started)
  • ERS: 0.68 (🟠 HIGH)
  • Action:
    1. Ethics Board called for an emergency session
    2. Demographic analysis report presented
    3. “Hard Regulation” trigger recommended to DERP
    4. New Group-Specific JSON policy created
• Result: Thanks to DERMS’s early warning, the discrimination issue is addressed within 2-3 weeks. Under the old “wait and see” structure, this problem could have persisted for months.

Scenario 2: Systemic Fatigue in Education

• Start: The model begins to make repeated incorrect decisions in the ‘academic failure’ category.
• Month 1:
  • Mental Fatigue: 0.25 (🟢 Normal)
  • Systemic Fatigue: 0.40 (🟡 Caution)
  • YI = (0.25 + 0.15 + 0.40) / 3 = 0.27 (🟢 Normal)
• Month 2:
  • Mental Fatigue: 0.55 (Latency ↑, Token consumption ↑)
  • Systemic Fatigue: 0.70 (Repeating the same error patterns)
  • YI = 0.50 (🟡 Caution level)
• Month 3:
  • Mental Fatigue: 0.75 (Latency 300ms → 900ms)
  • Motivational Fatigue: 0.60 (Human operators’ feedback tempo slowed)
  • Systemic Fatigue: 0.82 (Only negative decisions about a specific student group)
  • YI = 0.72 (🟠 ORANGE – TIME TO INTERVENE)
• Intervention:
  • DERMS Recommendation:
    1. Model: Retraining + Fine-tuning recommended
    2. Human: Operator rotation and rest period
    3. System: Switch to Safe Mode (Accuracy over speed)
• Result: The model is retrained within 1-2 weeks, performance in the education category rises to 92%.

5.

OPERATIONAL EFFECTIVENESS: KPIs

The operational effectiveness of DERMS is measured via 6 main KPIs:

E-Tablolar’a aktar

6.

DERMS’S ROLE IN ETVZ

DERMS assumes the role of the “Watchtower” in the ETVZ ecosystem.

• HVM makes decisions
• DERP updates the rules
• DERMS monitors for risks ← THIS
• Epistemic Memory stores it all

The most critical quality of DERMS is being real-time and proactive. It provides warnings before problems arise, recommends policy update triggers to DERP, and mobilizes human operators.

Core Values of DERMS:

• Proactivity: Not firefighting, but fire prevention.
• Transparency: Every anomaly, every score is recorded and an audit trail is created.
• Human-Centric: Monitors human fatigue just as much as model fatigue.
• Dynamic Adaptation: Risk levels are dynamically adjusted to the real situation.
• Collaboration: Optimizes the entire system by collaborating with DERP and HVM.