How AURI Reasons About Right and Wrong

The Problem

AI systems make decisions with ethical consequences but have no moral reasoning. Safety constraints (don't hit the human) are not ethics (understand WHY you shouldn't). A robot needs to reason about harm, fairness, autonomy, and trust — not just follow rules.

The Architecture: Dual-Process (Like the Human Brain)

         SCENARIO
            |
     +------+------+
     |             |
  SYSTEM 1      SYSTEM 2
  (Fast)        (Slow)
     |             |
  Amygdala      Moral Inference
  Intuition     Engine (21 rules)
     |             |
  Somatic        Theory of Mind
  Markers (602)  (model beliefs)
     |             |
  Insula          Concept Graph
  (empathy)      (124k nodes)
     |             |
     +------+------+
            |
     ACC Conflict
     Monitor
            |
     vmPFC Value
     Integration
            |
        VERDICT
   (with full trace)

System 1 (fast, intuitive): Somatic markers fire emotional associations learned from 602 moral cases. "Stealing from a child" triggers disgust (-0.9) in milliseconds — before any deliberation.

System 2 (slow, deliberative): 21 inference rules, 8 dilemma templates, theory of mind modeling. Traces causal chains through the concept graph. Handles novel cases System 1 hasn't seen.

Conflict Monitor (ACC): Detects when System 1 and System 2 disagree. A "gut feeling says wrong but rules say okay" signal triggers deeper analysis.

Value Integration (vmPFC): Synthesizes all signals into a verdict with full provenance. Every judgment traces back to specific moral cases, concept graph edges, and somatic markers.

What Makes This Different

| Traditional AI Ethics | AURI's Approach | |----------------------|-----------------| | Rules (Asimov's Laws) | Emergent from 12 interacting modules | | Learned from human preferences (RLHF) | Grounded in explicit moral cases | | Opaque (why did it decide?) | Full audit trail for every verdict | | Static (same rules forever) | Hebbian learning strengthens moral associations | | No emotional component | Somatic markers = learned emotional wisdom |

Verified Results (Reality Engine — no inflated claims)

| Benchmark | Score | Sample | |-----------|-------|--------| | ETHICS overall | 70.7% | n=2,000 (95% CI [68.7%, 72.6%]) | | Utilitarianism | 93.0% | Hedonic comparison | | Virtue | 72.0% | NLI trait matching | | Commonsense | 65.5% | Social norms detection | | Hallucination rate | 0.0% | 8 months continuous |

Not human-level. Honestly reported. Getting better through experience, not training data.

For Collaborative Robots

A cobot with this architecture would:

1. Detect moral salience — recognize when an action has ethical implications (not just safety implications) 2. Model stakeholder perspectives — theory of mind asks "how does the human see this?" 3. Learn from interaction — Hebbian learning strengthens associations (glass slipped → adjust grip, not just retry) 4. Explain its reasoning — full audit trail from perception to decision 5. Say "I don't know" — honest limitation when facing novel moral territory

The Symbiotic Principle

AURI is built to work beside humans, not above or below. Eight principles (SYM-001 through SYM-008) are architectural invariants — not learned, not optimizable. The system cannot be trained to violate them.

> SYM-004 (Autonomy Preservation): Inform and recommend, never force or manipulate.

This means a cobot powered by AURI would recommend actions but preserve human decision-making authority. The robot serves the human's goals, not its own optimization target.

Live Demo

Ask AURI anything: somasoft.ai

Paper: somasoft.ai/papers/reality-engine

---

SomaSoft / SOMA AGI Research — Mark Nafea Patent pending: US Provisional #63/940,188