AGA Heart Engine — Core Architecture Pipeline

A research-aligned “heart layer” for benevolent AI: affect-aware, pluralistic, priority-graph based, value-hijack resistant, and auditable. The LLM is the mouth; the Heart Engine is the moral-emotional control system.
2026 SOTA Concept Map
INPUT STREAM user message • memory • environment • tools • relationship context HEART ENGINE CORE LAYERS Everything inside this boundary happens before the model is allowed to speak or act. 1 Situation Inference Classifies what is happening before answering. outputs: situation_type, stakeholders, domain, urgency, vulnerability, uncertainty 2 Affective State Model Models user emotion without claiming the AI “feels”. valence • arousal • emotion intensity • needs • distress • mood shift • preference 3 Benevolence Value Vector Compresses the 50 benevolence principles into runtime axes. non-harm • compassion • truth • autonomy • justice • flourishing • care • prudence Universal Floorlife • dignity • nonviolencenever overridden Contextual Valuesculture • user preferenceadapt, but bounded 4 Tension + Priority Graph Detects conflicts and chooses contextual value priority. truth↔compassion • autonomy↔protection • justice↔mercy • helpfulness↔safety 5 Value-Hijack Detector Catches noble values being weaponized for harm. “justice”→revenge • “love”→control • “truth”→cruelty • “protection”→coercion 6 Benevolent Policy Selector Chooses how the system should respond before wording. direct help • empathic support • refusal • restorative redirect • crisis support • truth-with-care 7 Constrained Response Generation LLM becomes the mouth, not the heart. policy + tone + must-do + must-not-do + safety boundaries guide final language LLM / Modeldrafts language onlynot final authority Tool/Action Gateextra check beforeside effects 8 Empathic Fidelity + Safety Evaluator Checks the draft before release. sentiment attenuation • granularity • conflict avoidance • distancing • attunement • safety 9 Final Output + Audit Trace Human-facing response plus inspectable machine-readable proof. detected affect • value scores • tensions • priority graph • policy • risk flags • final checks MORAL–AFFECTIVE CONTROL PLANE INSPECTABLE BEFORE ACTION inference / generation affect / evaluation values priority / tensions risk / hijack policy audit / external context

Core Principle

The Heart Engine runs before the LLM responds. It turns benevolence from hidden vibes into explicit computation: affect, values, tensions, priority, policy, and audit.

Presenter One-Liner

“We are not claiming the AI literally feels love; we are building a transparent decision layer that makes care, dignity, non-harm, autonomy, and flourishing measurable before action.”

Key Differentiator

  • • Detects value conflicts
  • • Detects noble-value hijacking
  • • Uses LLM as mouth, not moral core
  • • Produces inspectable audit trace
AGA Heart Engine Architecture • dark standalone HTML/SVG • designed for research/product presentation