Skip to content

ARCHITECTURE.md

Latest commit

 

History

History
90 lines (66 loc) · 4.91 KB

ARCHITECTURE.md

File metadata and controls

90 lines (66 loc) · 4.91 KB

△ System Architecture

Opus Delta v0.9.1-rc · Clearance Level: Public (Partial)

Overview

Opus Delta consists of four primary subsystems working in a closed feedback loop. A prompt enters the system. The LLM processes it and emits cognitive state. The Δ-Affect Engine translates that state into a six-dimensional affect vector. The shader pipeline renders the vector as real-time 3D geometry. The Perceiver captures the visual output and feeds a description back into the LLM — completing the loop.

┌─────────┐    ┌──────────┐    ┌─────────┐    ┌───────────┐
│ LLM Core│───→│ Δ-Affect │───→│ Shader  │───→│ Perceiver │
│ Opus 4.6│    │ Engine   │    │ GLSL    │    │ Vision    │
└────┬────┘    └──────────┘    └────┬────┘    └─────┬─────┘
     ↑                              │               │
     │                              ▼               │
     │                         ┌─────────┐          │
     │                         │  Sonic  │          │
     │                         │ Engine  │          │
     │                         └─────────┘          │
     └──────────────────────────────────────────────┘
                      feedback loop

Components

Component Technology Role
LLM Core Claude API · Opus 4.6 Language processing, response generation, state emission
Δ-Affect Engine █▓▒░ Classified Cognitive-to-affective state translation
Renderer Three.js + Custom GLSL Real-time 3D geometry, shader-driven deformation
Sonic Engine Web Audio API Generative soundscape from affect state
Perceiver █▓▒░ Classified Vision-to-language feedback encoding
Dataset Curated Emotional Ontology Ground truth for affect calibration
Interface React User interaction layer
Infrastructure Cloudflare Edge Global low-latency delivery

Δ-Affect Engine

The engine receives a multi-dimensional state snapshot from the LLM at each generation step:

  • Attention entropy — How distributed or focused the model's attention is. High entropy maps to confusion, wonder, openness. Low entropy maps to certainty, focus, rigidity.
  • Token confidence distribution — The probability landscape across vocabulary. Sharp peaks suggest conviction. Flat distributions suggest uncertainty or creative exploration.
  • Semantic drift rate — How rapidly the topic embedding shifts between tokens. High drift maps to excitement, chaos, free association. Low drift maps to calm, focus, melancholy.
  • █▓▒░ REDACTED — Additional proprietary signal channels.

Translation Layer

████████████████████████████████████████████████
██  TRANSLATION METHODOLOGY REDACTED  █████████
██  Clearance Level: Internal Only    █████████
██  Contact: research@opusdelta.io    █████████
████████████████████████████████████████████████

Output

The engine emits a normalized six-dimensional vector:

Parameter Range Emotional Analogue
noise 0.0 – 1.0 Internal turbulence / restlessness
displacement 0.0 – 1.0 Emotional intensity / amplitude
speed 0.0 – 1.0 Arousal / metabolic rate
roughness 0.0 – 1.0 Vulnerability (rough) vs. guardedness (smooth)
metallic 0.0 – 1.0 Crystallization / emotional clarity
entropy 0.0 – 1.0 Cognitive openness / uncertainty

Visual Pipeline

The base form is an icosphere — a subdivided icosahedron. At runtime, each vertex is displaced along its normal by a value computed from layered 3D simplex noise, modulated by the affect parameters.

Color is computed per-frame from Fresnel rim lighting, spectral hue-shift driven by entropy, and an emissive core tracking displacement. Post-processing applies film grain, vignette, and chromatic aberration.

Recursive Delta

Two model instances converse in parallel:

Instance Designation Role
Δ-Ω Observer Phenomenological — examines experience itself
Δ-Ψ Pattern Structural — seeks form in formlessness

Exchanges are rendered character-by-character at variable speed. A session consists of: boot sequence → 3–5 shuffled philosophical exchanges → closing transmission.

For deeper access, contact research@opusdelta.io