Trust Dynamics

Formalizing Trust and Reputation Evolution in Strategic Coopetition (TR-2)

This document provides a comprehensive treatment of trust dynamics from Technical Report 2 (TR-2025-02), explaining how trust evolves through repeated interactions and how it affects strategic behavior.

Executive Summary

For Practitioners: Trust is not static, it evolves based on observed behavior. A single betrayal can destroy years of trust-building, but consistent cooperation slowly rebuilds confidence. Understanding these dynamics is essential for managing long-term partnerships.

For Researchers: We formalize trust as a two-layer dynamic system: immediate trust (T) responds to current behavior while reputation damage (R) tracks violation history. Asymmetric updating (3:1 negativity bias) and trust ceiling mechanisms create path-dependent dynamics validated against the Renault-Nissan Alliance.

Conceptual Foundation

Why Trust Matters in Coopetition

In coopetitive relationships, actors face ongoing temptation to defect, to capture short-term gains at the partner’s expense. Trust addresses this by: 1. Enabling Cooperation: High trust reduces perceived exploitation risk

  1. Gating Information Sharing: Actors share more with trusted partners
  2. Supporting Long-Horizon Planning: Trust enables commitment to joint investments
  3. Creating Relationship Value: Trusted partnerships are more productive

The Trust Evolution Challenge

Trust is inherently dynamic:

Classical game theory lacks mechanisms for this dynamic evolution. Our formalization addresses this gap.

Distinguishing Trust Concepts

Concept Definition Our Formalization
Immediate Trust Current confidence in partner $T_{ij} \in [0, 1]$
Reputation Historical track record $R_{ij} \in [0, 1]$ (damage)
Trustworthiness Actual reliability Not modeled (observable only through signals)
Trust Ceiling Maximum achievable trust $\Theta = 1 - R$

Two-Layer Trust Architecture

Layer 1: Immediate Trust ($T_{ij}$)

Definition: Actor $i$’s current confidence in actor $j$’s reliability and cooperative intent.

Trust Evolution Curve

Properties:

Interpretation:

T Value Interpretation Behavioral Implication
0.0 - 0.2 No trust Expect defection; protect self
0.2 - 0.4 Low trust Cautious engagement; verify behavior
0.4 - 0.6 Moderate trust Willing to cooperate with monitoring
0.6 - 0.8 High trust Confident cooperation; share information
0.8 - 1.0 Full trust Deep partnership; strategic alignment

Layer 2: Reputation Damage ($R_{ij}$)

Definition: Accumulated history of actor $j$’s violations from actor $i$’s perspective.

Properties:

Interpretation:

R Value Interpretation Trust Ceiling (Θ)
0.00 No violation history 1.00 (full recovery possible)
0.20 Minor past issues 0.80
0.40 Significant violations 0.60
0.60 Severe damage 0.40
0.80 Major betrayals 0.20
1.00 Complete destruction 0.00 (trust impossible)

The Trust Ceiling Mechanism

Equation 7 (TR-2):

\[\Large \Theta_{ij}^t = 1 - R_{ij}^t\]

Trust Ceiling

The trust ceiling creates hysteresis: even perfect cooperation cannot restore trust beyond this limit.

Example Trajectory:

Initial state: T=0.50, R=0.00, Θ=1.00
After violation: T=0.30, R=0.40, Θ=0.60
After 10 periods cooperation: T→0.58 (cannot exceed 0.60)
After 50 periods cooperation: T→0.60 max (ceiling reached)

The relationship can improve but will always bear the scar of past violations.

Cooperation Signal

Assessing Partner Behavior

To update trust, actors assess whether partners are cooperating or defecting. This requires comparing observed behavior to expectations.

Equation 4 (TR-2): Cooperation Signal

\[\Large s_{ij}^t = \tanh\left(\kappa \cdot (a_j^t - a_j^{\text{baseline}})\right)\]

Components:

Component Meaning Typical Value
$a_j$ Actor $j$’s actual action Observed cooperation level
$a_j^{\text{baseline}}$ Expected cooperation Context-dependent norm
$\kappa$ Signal sensitivity 1.0 (default)
$s_{ij}$ Cooperation signal $\in (-1, 1)$

Signal Interpretation

s Value Interpretation Trust Effect
s > 0.5 Strong positive signal Rapid trust building
0 < s < 0.5 Mild cooperation Gradual trust building
s ≈ 0 Met expectations Neutral (slight positive)
-0.5 < s < 0 Mild defection Gradual trust erosion
s < -0.5 Strong violation Rapid trust erosion

Why Bounded (tanh)?

The hyperbolic tangent function ensures:

  1. Bounded signals: Even extreme deviations produce finite trust changes
  2. Smooth transitions: No discontinuities in trust dynamics
  3. Diminishing sensitivity: Very large deviations don’t dominate

Baseline Elicitation

The baseline $a_j^{\text{baseline}}$ is context-specific:

Context Baseline Determination
Contractual Agreed-upon cooperation levels
Historical Moving average of past behavior
Normative Industry/organizational standards
Rational Nash equilibrium prediction

Asymmetric Trust Evolution

The Negativity Bias

Trust evolution is fundamentally asymmetric: violations hurt more than cooperation helps.

Negativity Bias

Equation 5 (TR-2): Trust Change

\[\Delta T_{ij}^t = \begin{cases} \lambda^+ \cdot s_{ij}^t \cdot (\Theta_{ij}^t - T_{ij}^t) & \text{if } s_{ij}^t > 0 \text{ (building)} \\ -\lambda^- \cdot |s_{ij}^t| \cdot T_{ij}^t \cdot (1 + \xi \cdot D_{ij}) & \text{if } s_{ij}^t \leq 0 \text{ (erosion)} \end{cases}\]

Trust Building ($s > 0$)

\[\Large \Delta T = \lambda^+ \cdot s \cdot (\Theta - T)\]

Components:

Properties:

Trust Erosion ($s \leq 0$)

\[\Large \Delta T = -\lambda^- \cdot |s| \cdot T \cdot (1 + \xi \cdot D_{ij})\]

Components:

Properties:

The 3:1 Ratio

Validated Finding: Trust erodes approximately 3× faster than it builds.

\[\Large \text{Negativity Ratio} = \frac{\lambda^-}{\lambda^+} = \frac{0.30}{0.10} = 3.0\]

Empirical Grounding: This ratio aligns with behavioral economics research on negativity bias (Rozin & Royzman, 2001; Slovic, 1993):

Interdependence Amplification

Equation Component: $(1 + \xi \cdot D_{ij})$

When actor $i$ depends heavily on actor $j$ (high $D_{ij}$), violations by $j$ cause amplified trust damage:

\[\Large \text{Amplification factor} = 1 + \xi \cdot D_{ij}\]
Dependency Level $D_{ij}$ Amplification Factor
Low dependency 0.2 1.10
High dependency 0.8 1.40

Intuition: Betrayal by someone you depend on hurts more because:

  1. Direct goal impact (structural coupling)
  2. Limited alternatives (can’t easily switch)
  3. Psychological salience (greater attention to critical dependencies)

Benchmark Evidence: High-dependency relationships experience 27% faster trust erosion for equivalent violations.

Reputation Dynamics

Reputation Damage Accumulation

Equation 8 (TR-2): Reputation Change

\[\Large R_{ij}^{t+1} = R_{ij}^t + \Delta R_{ij}^t - \delta_R \cdot R_{ij}^t\]

Damage Term:

\[\Delta R_{ij}^t = \begin{cases} \mu_R \cdot |s_{ij}^t| \cdot (1 - R_{ij}^t) & \text{if } s_{ij}^t < 0 \text{ (violation)} \\ 0 & \text{if } s_{ij}^t \geq 0 \text{ (no damage)} \end{cases}\]

Parameters

Parameter Symbol Validated Value Meaning
Damage Severity $\mu_R$ 0.60 How much damage per violation
Decay Rate $\delta_R$ 0.03 Forgetting rate per period

Damage Mechanics

Forgetting Dynamics

Reputation decays slowly even without new violations:

\[\Large R(t) = R_0 \cdot (1 - \delta_R)^t\]

Time to decay from $R=0.50$ to $R=0.25$ (half-life):

With $\delta_R = 0.03$, forgetting is slow, violations leave lasting marks.

Trust-Performance Relationship

Benchmark Validation

From 760 experiments across 20 algorithms:

Trust-Return Correlation: r = 0.552 (p < 0.001)

Trust-Return Scatter Plot

Algorithm Performance by Trust

Trust Category Algorithm Examples Mean Return
High Trust (>0.8) Constant_050, ISAC 72,494
Medium Trust (0.4-0.8) MADDPG, VDN 50,903
Low Trust (<0.4) IPPO, SelfPlay_PPO 21,019

The Constant_050 vs Random Natural Experiment

Metric Random Constant_050 Insight
Mean Cooperation 47.7% 55.0% +7.3%
Cooperation Variance ~29% 0% -29%
Final Trust 7.1% 98.7% +91.6%
Mean Return 24,939 72,494 +190%

Critical Insight: Constant_050’s superior performance comes from consistency, not higher average cooperation. Unpredictable cooperation triggers negativity bias, eroding trust despite moderate average behavior.

Hysteresis and Path Dependence

What is Hysteresis?

Hysteresis means the system’s state depends on its history, not just current inputs. In trust:

Formal Characterization

Trust Ceiling:

\[\Large \Theta(t) = 1 - R(t)\]

For any trajectory with violations:

\[\Large \max[T(t \to \infty)] = \Theta < 1.0\]

Even infinite cooperation cannot restore trust to pre-violation levels.

Recovery Dynamics

Validated Finding (TR-2 §7.3): Median recovery ratio = 1.11

After 35 periods of sustained cooperation following a violation:

Practical Implications

  1. Prevention over cure: Avoiding violations is more efficient than recovering
  2. Reputation management: Long-term reputation is a strategic asset
  3. Crisis planning: Recovery requires sustained effort over extended periods
  4. Relationship investment: High-trust relationships are valuable and fragile

Trust-Gated Reciprocity

Extending Utility Functions

Trust modulates cooperative behavior through utility augmentation:

Equation (TR-2): Trust-Gated Reciprocity

\[\Large U_i(\mathbf{a}, \mathbf{T}^t) = U_i^{\text{base}}(\mathbf{a}) + \sum_{j \neq i} \rho \cdot T_{ij}^t \cdot (a_j - a_j^{\text{baseline}}) \cdot a_i\]

Components:

Mechanism

When trust is high ($T_{ij} \to 1$):

When trust is low ($T_{ij} \to 0$):

Strategic Implications

Trust gates cooperation even when mutual benefit is objectively available:

Scenario T Level Cooperation? Rationale
High T, partner cooperates High Yes Reciprocity rewarded
High T, partner defects High Maybe Still some trust
Low T, partner cooperates Low No Don’t believe it will continue
Low T, partner defects Low No Expect exploitation

Validation: Renault-Nissan Alliance

Case Background

The Renault-Nissan Alliance (1999-present) provides a longitudinal test case with documented trust evolution across five phases:

Phase Period Trust Dynamic
Crisis 1999-2001 Low trust, high stakes
Recovery 2001-2005 Trust building through results
Growth 2005-2010 Expanding cooperation
Ghosn Era 2010-2018 Centralized trust in leader
Post-Ghosn 2018-2025 Crisis and partial recovery

Model Calibration

renault_nissan_params = {
    'n_agents': 2,
    'max_steps': 26,      # 26 years
    'lambda_plus': 0.08,  # Cross-cultural: slower building
    'lambda_minus': 0.32, # Standard erosion
    'mu_R': 0.65,         # Visible scandals: higher damage
    'delta_R': 0.02,      # Slower forgetting
    'T_init': 0.30,       # Started in crisis
    'R_init': 0.20,       # Some initial damage
}

Validation Results

Overall Score: 49/60 (81.7%)

Phase Predicted Pattern Historical Match Score
Crisis Low trust, cautious cooperation ✓ 10/12
Recovery Gradual trust building ✓ 10/12
Growth High cooperation, trust peak ✓ 10/12
Ghosn Era Stable high trust Partial 9/12
Post-Ghosn Sharp decline, partial recovery ✓ 10/12

Key Dynamics Captured

  1. Crisis Response: Low initial trust matched historical caution
  2. Recovery Path: Gradual trust building through consistent results
  3. Ghosn Scandal Impact: Sharp trust erosion matching documented crisis
  4. Hysteresis: Post-recovery trust ceiling below pre-scandal levels

Implementation Details

Code Correspondence

Trust dynamics are implemented in coopetition_gym/core/trust_dynamics.py:

def compute_cooperation_signal(action, baseline, kappa=1.0):
    """Equation 4: Bounded cooperation signal."""
    return np.tanh(kappa * (action - baseline))

def update_trust(T, signal, ceiling, lambda_plus, lambda_minus,
                 xi, D_ij):
    """Equation 5: Asymmetric trust evolution."""
    if signal > 0:
        # Trust building
        delta_T = lambda_plus * signal * (ceiling - T)
    else:
        # Trust erosion with interdependence amplification
        delta_T = -lambda_minus * abs(signal) * T * (1 + xi * D_ij)
    return np.clip(T + delta_T, 0.0, 1.0)

def update_reputation(R, signal, mu_R, delta_R):
    """Equation 8: Reputation damage evolution."""
    if signal < 0: delta_R_damage = mu_R * abs(signal) * (1 - R)
    else: delta_R_damage = 0.0
    decay = delta_R * R
    return np.clip(R + delta_R_damage - decay, 0.0, 1.0)

def compute_trust_ceiling(R):
    """Equation 7: Trust ceiling from reputation."""
    return 1.0 - R

Environment Integration

import coopetition_gym

# Create environment with custom trust parameters
env = coopetition_gym.make("TrustDilemma-v0",
    lambda_plus=0.10,
    lambda_minus=0.30,
    mu_R=0.60,
    delta_R=0.03,
    xi=0.50,
    kappa=1.0,
    T_init=0.50,
    R_init=0.00,
)

# Access trust state
obs, info = env.reset()
trust_matrix = info['trust_matrix']
reputation_matrix = info['reputation_matrix']

Practical Applications

For Partnership Management

  1. Monitor trust trajectory: Track T over time for early warning
  2. Prevent violations: Small violations have outsized impact
  3. Plan for recovery: Budget extended cooperation after incidents
  4. Manage reputation: R damage is persistent; protect reputation

For Algorithm Design

  1. Consistency over magnitude: Predictable cooperation builds trust
  2. Avoid defection spirals: Single defection can be catastrophic
  3. Long horizon: Trust benefits compound over time
  4. Trust as state: Include trust in policy state representation

For Environment Customization

# Post-crisis scenario
crisis_env = coopetition_gym.make("TrustDilemma-v0",
    T_init=0.20,   # Low starting trust
    R_init=0.50,   # Significant reputation damage
)

# High-trust scenario
trusted_env = coopetition_gym.make("TrustDilemma-v0",
    T_init=0.80,   # High starting trust
    R_init=0.00,   # No reputation damage
)

Further Reading

Primary Source

Background

Benchmark Analysis

Technical Reports