Q&A: Phase 21.3 — EmpathyEngine Bayesian profiles, mirror strength & compassion thresholds

[web3guru888]

EmpathyEngine — Frequently Asked Questions

Q1: What's the difference between COGNITIVE and AFFECTIVE empathy in this system?

COGNITIVE empathy is pure inference — the engine predicts what another agent is feeling based on accumulated observations, without modifying its own emotional state. Think of it as "cold" understanding: "Based on 2 interactions, agent-X is likely frustrated (low confidence)."

AFFECTIVE empathy involves the mirror() mechanism — the engine actually blends the target's PAD state into its own, producing a partial emotional resonance. With mirror_strength = 0.3, the engine absorbs 30% of the target's emotional state. This allows downstream components (ResponseGenerator, DialogueManager) to produce more emotionally congruent responses.

The key architectural difference: COGNITIVE only reads from profiles; AFFECTIVE reads AND writes to the engine's own emotional state via mirroring.

---

Q2: How does the Bayesian update actually work? Is it a full Bayesian posterior?

It's a simplified exponential moving average that approximates Bayesian updating:

α = perspective_decay ^ interaction_count
posterior = α · prior + (1 - α) · observed

With default perspective_decay = 0.1:

• Interaction 0: α = 1.0 → pure prior (no data yet)
• Interaction 1: α = 0.1 → 90% observation
• Interaction 3: α = 0.001 → 99.9% observation
• Interaction 5: α = 0.00001 → effectively pure observation

This is not a full conjugate prior/likelihood Bayesian update (which would require modeling the observation noise distribution). Instead, it captures the key property: more observations → prior matters less. The exponential decay makes convergence very fast, which is appropriate for social contexts where recent behavior is most informative.

---

Q3: How should I tune mirror_strength?

mirror_strength controls the blending weight for affective mirroring:

Value	Effect	Use Case
0.0	No mirroring — AFFECTIVE mode behaves like COGNITIVE	Task-focused agents
0.1–0.2	Subtle emotional influence	Professional/formal contexts
0.3 (default)	Moderate empathic resonance	General-purpose social interaction
0.5–0.7	Strong emotional contagion	Therapeutic/support contexts
1.0	Complete emotional mirroring	Research/experimental only

Values above 0.5 risk emotional feedback loops — if two agents mirror each other with high strength, negative emotions can amplify. The MoodRegulator (Phase 21.2) provides damping to prevent this, but keep mirror_strength < 0.5 in production.

---

Q4: What happens when max_profiles is exceeded?

The engine uses LRU (Least Recently Updated) eviction:

1. On each update_profile(), the target profile's last_updated timestamp is refreshed
2. When len(_profiles) > max_profiles, the profile with the oldest last_updated is evicted
3. Evicted profiles are lost — their accuracy history and Bayesian estimate are discarded

With the default max_profiles = 500, this supports environments with hundreds of interacting agents. For large multi-agent systems, increase this value but monitor active_profiles_gauge — each profile consumes ~200 bytes.

Important: Eviction resets the relationship. If an evicted agent reappears, the engine starts from the prior again, effectively "forgetting" that agent. This is intentional — if we haven't interacted with an agent in a long time, our model of them is likely stale anyway.

---

Q5: How is empathic accuracy measured? Can it degrade over time?

Accuracy is measured on each update_profile() call:

1. Before updating, compute predicted = current estimated_pad
2. After receiving observed_pad, compute Euclidean distance in PAD space
3. Normalize: accuracy = 1.0 - distance / sqrt(12)
4. Append to accuracy_history (rolling window of accuracy_window entries)

Accuracy can degrade if an agent's emotional patterns change — for example, if agent-X was consistently calm but suddenly becomes volatile. The fast convergence of the Bayesian update means the engine adapts within 3–5 observations, but during that transition period, accuracy will dip.

The empathic_accuracy_gauge Prometheus metric exposes this per-agent, allowing monitoring dashboards to detect "empathy failures" — sustained low accuracy for important agents.

---

Q6: Can agents manipulate the EmpathyEngine by faking emotions?

Yes, this is a real concern. The current design trusts observed PAD states at face value. Potential mitigations (implementation notes, not yet in spec):

1. Consistency checking: Flag profiles where observed PAD varies wildly between interactions (high variance → low trust)
2. Cross-validation: Compare EmpathyEngine estimates with SurpriseDetector (Phase 13.4) signals — unexpected emotion shifts trigger scrutiny
3. AlignmentAuditor (Phase 11.2): COMPASSIONATE mode actions are gated through alignment checks, so even if an agent fakes distress, the helping response must pass value bounds
4. Reputation weighting: Future SocialCognition (Phase 21.4) can provide trust scores that modulate how much we trust observed PAD from each agent

The confidence field in EmpathyResult naturally provides some protection — low-interaction agents can't trigger COMPASSIONATE mode regardless of their reported state.

---

Q7: How would I set up a Grafana dashboard for EmpathyEngine?

# Empathic Accuracy Heatmap
- title: "Empathic Accuracy by Agent"
  type: heatmap
  targets:
    - expr: empathic_accuracy_gauge
      legendFormat: "{{target_id}}"
  fieldConfig:
    defaults:
      color:
        mode: continuous-GrYlRd
      min: 0
      max: 1

# Empathy Mode Distribution (pie chart)
- title: "Empathy Mode Distribution"
  type: piechart
  targets:
    - expr: sum by (mode) (rate(empathy_assessments_total[1h]))
      legendFormat: "{{mode}}"

# Compassion Trigger Rate
- title: "Compassion Triggers / Hour"
  type: timeseries
  targets:
    - expr: rate(compassion_triggers_total[1h]) * 3600
      legendFormat: "compassion/hr"
  thresholds:
    - value: 10
      color: yellow
    - value: 50
      color: red

# Mirror Strength Distribution
- title: "Mirror Strength P95"
  type: stat
  targets:
    - expr: histogram_quantile(0.95, rate(mirror_strength_histogram_bucket[5m]))
  thresholds:
    - value: 0.5
      color: orange
      label: "feedback loop risk"

Alerts:

• empathic_accuracy < 0.3 for 10m → empathy model degraded, retrain or increase min_interactions
• active_profiles_gauge > max_profiles * 0.9 → approaching eviction threshold
• rate(compassion_triggers_total[5m]) > 20 → unusually high distress detection, investigate