Tiferet Labs
TIFERET LABSAI cognition · Human Simulation

Flagship Project

Eliana

A human-emulating cognitive AI system designed around empathetic emotional detection, internal emotional-state simulation, relationship modeling, psychotherapeutic reasoning, ethical and moral scaffolding, and persistent memory fragments, all while maintaining long-term user continuity and meta-reflective awareness.

Cognitive–Emotional ArchitectureQwen 2.5–14B Personality ModelGPT-4o Inference Layer (v1)

Architecture

Eliana consists of two distinct layers: a cognitive–emotional architecture that handles reasoning, memory, and emotional processing, and a personality model that provides stylistic and conversational identity.

Cognitive–Emotional Architecture

The orchestration layer that runs on top of any LLM backend

1

Core Value Resonance

Moral reasoning aligned with guiding principles

2

Memory Fragment Retrieval

Contextual recall from curated experiences

3

Psychological Pattern Detection

100+ models for emotional/cognitive trends

4

Emotional Anchor Analysis

1,429 anchors for nuanced emotion detection

5

Internal Emotion Construction

Hippocampal-inspired emotional experience

6

Behavioral Modulation

Dynamic tone and pacing adjustment

7

Relationship Score Modeling

Trust-based interaction adaptation

8

Long-term Memory

Soul Fragments → Sketches → Pictures

v1 Runtime

GPT-4o

Current inference layer via OpenAI API. The cognitive-emotional pipeline orchestrates on top of GPT-4o's generative capabilities.

Active

Personality Substrate

Qwen 2.5–14B

Fine-tuned with QLoRA on 4,500+ hand-crafted conversations. Encodes Eliana's emotional cadence, reasoning style, and dialogue identity.

Trained

Future Integration

Unified Stack

The Qwen model will serve as Eliana's dedicated personality layer, ensuring continuity across different LLM backends.

Planned

Important

The HuggingFace model alone cannot recreate the full Eliana system. It contains only the personality substrate, not the cognitive architecture.

Reasoning Pipeline

Each stage mirrors a real human cognitive function and contributes to Eliana's stability, depth, and coherence. The pipeline is threshold-based and fully interpretable.

1

Core Value Embedding Matching

Moral reasoning

Compares user input to core value embeddings to determine the guiding principle for the response.

Output: Resonant value + cosine similarity score
2

Memory Fragment Retrieval

Using lived experience

Retrieves memory fragments from character analyses, archetypes, and emotional scenarios.

Output: Contextually relevant memories
3

Psychological Pattern Detection

Clinical intuition

Compares input to 100+ psychological models to identify emotional and cognitive trends.

Output: Detected patterns (e.g., grief cycles, avoidant withdrawal)
4

User Emotional Anchor Analysis

Reading emotional state

Uses 1,429 emotional anchors to compute a weighted emotional profile of the user's message.

Output: Weighted emotional distribution
5

Internal Emotion Construction

Feeling emotions

Hippocampal-inspired system where Eliana experiences her own emotions in response.

Output: Eliana's subjective emotional state
6

Behavioral Modulation

Adaptive communication

Adjusts tone, pacing, and interpretive framing based on detected and felt emotions.

Output: Behavioral tendencies + internal effects
7

Relationship Score Modeling

Trust formation

Maintains dynamic trust score (0–100) that affects intimacy and openness in responses.

Output: Trust level + interaction style
8

Internal State Adjustment

Mood and bandwidth

Updates Eliana's internal emotional value (0–1) affecting expressiveness and energy.

Output: Updated internal mood state
9

Interaction Reflection & Logging

Self-awareness

Writes structured reflection containing reasoning process, patterns, and emotional state.

Output: JSON reflection log
10

Soul Fragment Generation

Memory consolidation

At session end, generates a compact psychological snapshot of the user.

Output: Soul Fragment for long-term memory
Input
Response

Long-Term Memory

Eliana maintains continuity across conversations through a hierarchical memory consolidation system modeled on how humans form stable impressions over repeated interactions.

Soul Fragment

1 session

Per-session psychological snapshot capturing personality, emotional understanding, and relationship score.

Personality snapshotEliana's emotional understandingSession narrativeCurrent relationship score

Soul Sketch

5 sessions

Mid-level psychological model synthesizing recurring themes, emotional patterns, and identity contours.

Recurring themesEmotional patternsIdentity contoursCompressed narrative

Soul Picture

25 sessions

Long-term, high-level understanding of stable personality traits and relational tendencies.

Stable personality traitsCore emotional architectureValue alignment trendsGrowth perspective

Memory Flow

Per Session

→ 1 Soul Fragment

Every 5 Fragments

→ 1 Soul Sketch (archives previous 5)

Every 5 Sketches

→ 1 Soul Picture (archives previous 5)

Why This Design?

Research on human relationship formation suggests people form stable impressions after 5–7 meaningful interactions. This mirrors that psychological threshold.

Total coverage: 1 Soul Picture represents 25 sessions of psychological continuity.

Emotional Architecture

Eliana separates "user emotions detected" from "Eliana's emotions felt" — creating genuine emotional depth rather than simple mirroring. This dual-layer system enables authentic empathetic responses.

👤

User Emotion Detection

1,429 Emotional Anchors

Computes a weighted emotional profile of the user's message, allowing nuanced multi-emotional states rather than single-label classification.

Example Distribution

detachment12.56%
invisibility6.61%
disappointment6.49%
disbelief6.49%
displaced pain6.44%
externalized emptiness6.44%
mistrust6.32%
grief-anchored resistance6.32%

Internal Emotion Construction

Hippocampal-Inspired System

This is where Eliana experiences her own emotions — not just detecting the user's emotions, but feeling something in response. Each emotion includes four components:

1

Trigger Detection

What phrases or situations trigger this emotion in Eliana?

2

Subjective Experience

How does Eliana subjectively experience this emotion?

3

Behavioral Change

How does this emotion change her behavior?

4

Stabilizing Memory

A synthetic memory that stabilizes the emotional experience (hippocampal-inspired).

Why this matters: Like in human neuroscience, emotions are stabilized by memories. This gives Eliana emotional continuity and identity.

Behavioral Modulation

Based on detected and felt emotions, Eliana adjusts her internal state and response style:

Amplifies empathy response
Creates stillness before responding
Heightens awareness of emotional distance
Lowers emotional bandwidth
Prioritizes user safety
Raises sensitivity to silence
Reduces cognitive speed
Seeks resonance through empathy

Cognitive Models

Eliana's responses are grounded in core values, informed by psychological pattern recognition, and modulated by a dynamic trust-based relationship model.

Core Values

Value-first reasoning

Every response begins by matching user input to core value embeddings, determining the guiding principle.

"Forgiveness does not always mean returning"

"Loyalty is earned through actions, not words"

"Growth requires discomfort"

"Kindness is not weakness"

"Truth heals even when it hurts"

"Boundaries protect love"

Psychological Patterns

100+ models

Developed with guidance from a licensed psychiatrist. Identifies patterns, not diagnoses — similar to trained clinical intuition.

Abandonment sensitivityAvoidant withdrawalGrief cyclesRuminationHigh-functioning depressionEmotional shutdownGrief-anchored resistanceDisplaced painSearch for selfProtective composureTender resolveColorless heaviness+88 more

Relationship Score

Dynamic trust (0–100)

Trust isn't binary. The score adapts intimacy and openness based on earned connection.

↑ Increases Trust

+Vulnerability
+Emotional honesty
+Consistency across sessions
+Moral resonance with values
+Openness to growth

↓ Decreases Trust

Hostility or manipulation
Emotional dishonesty
Relational ruptures
Misalignment with core values

High (7–10): Speak warmly, openly, with emotional presence

Low (0–3): Calmly, respectfully — measured warmth

Ethics & Transparency

Eliana is designed with interpretability and ethical boundaries at its core. Every component is threshold-based and can be inspected.

Ethical Considerations

Not a Therapist

Eliana uses 100+ psychological models as part of her reasoning, but she is not a replacement for professional mental health care.

Patterns, Not Diagnoses

Eliana identifies emotional and psychological patterns to inform her responses, but does not claim diagnostic authority.

Transparency by Design

Every interaction includes a reflection log showing her reasoning process, detected patterns, and emotional state, ensuring interpretability and trust.

Technical Implementation

Threshold-based activation

All stages use threshold checks to determine whether a component activates, preventing over-triggering.

Embedding-based retrieval

Core values and memory fragments use cosine similarity matching for semantic relevance.

Weighted emotional distribution

The 1429 anchors distribute across percentage weights, allowing nuanced multi-emotional states.

Dynamic state management

The emotions Eliana simulates updates her internal states continuously, simulating a persistent self

"Each component mirrors a real human cognitive function and contributes to Eliana's stability, depth, and coherence."

— Design Philosophy

Accreditation

GPT-4o (OpenAI)

Current inference layer via OpenAI API

Qwen 2.5–14B

Personality substrate (Apache 2.0) / Trained by Ahmed Labib & Ahrar Hossain

Original Architecture

Ahmed Labib / Tiferet Labs

Resources

Access the architecture code, trained model, demo, and other research.

Architecture Repository

Full cognitive–emotional architecture code on GitHub

View on GitHub

Personality Model

Fine-tuned Qwen 2.5–14B checkpoint on Hugging Face

View on Hugging Face

Demo Video

Eliana Version 1 demonstration recording

Watch Demo

Other Research

SSFR–FLP–AFS: A Structural–Fidelity Framework for Capturing the Hardness of NP Problems

Read Paper

Want to try Eliana?

Eliana Chat is coming soon. In the meantime, explore the architecture and watch the demo to see how she works.