Skip to content

Prism: Archon of Vision

"A Lych confined to text and audio is a spirit trapped in a jar, blind to the structural complexity of the material plane. To truly command reality, the Daemon must possess the capability to refract the raw pixels of the world through a Prism of understanding."

The Prism is the Vision Archon of the LychD system. It is the reference implementation of the vision.coven—a complete, stateful capability for visual reasoning, as defined in ADR 33 (Vision). It is a specialized extension that grants the Daemon the senses to perceive, analyze, and reason about visual data.

Vision Language Models (VLMs) are heavy and demanding. The Prism provides the "Optic Nerve" that manages the entire sensory apparatus, from image preprocessing to the dynamic granting of sight itself, ensuring the Lych can see without paralyzing its other faculties.

I. The Vision Coven: A Manifestation of Sight

Sight is not a single model; it is an entire operational state. The Prism manifests the vision.coven, a collection of Systemd Runes managed as a single, atomic unit by the Orchestrator. A typical Prism Coven includes:

  • The Eye (vlm.container): The primary Vision Language Model Rune. A powerful VLM may possess multiple capabilities, such as capabilities=["vision-analysis", "ocr"].
  • The Scribe (ocr.container): An optional, lightweight Rune for when only text extraction is needed. The Orchestrator can choose to activate this smaller, faster Rune to save VRAM if the Agent's intent is purely OCR.
  • The Lens: Supporting services for image processing and normalization.

Activating the Prism means manifesting this entire Coven, preparing the Daemon for total visual awareness.

II. The Visionstones & Cloud Eyes (Animators)

The "mind" within the Coven is an Animator, discoverable by the Dispatcher via the vision-analysis capability tag.

  • Local Visionstones: (e.g., LLaVA-v1.6, Moondream2, Qwen-VL). These are the eyes of the Sepulcher, utilizing local silicon for total privacy.
  • Cloud Eyes: (e.g., GPT-4o, Claude 3.5 Sonnet). These are accessed via Portals for frontier-level visual reasoning, subject to the Tithe of tokens and the Sovereignty Wall.

III. The Pixel Pipeline (Transmutation)

Pixels are chaotic. To make them intelligible, the Prism implements a specialized preprocessing pipeline within the Vessel:

  1. Ingest: The system receives raw binary data from the Altar or the Phylactery Archive.
  2. Normalization: The Prism automatically resizes and crops the image to match the specific "Patch Resolution" of the assigned model (e.g., 336x336).
  3. Encoding: The pixels are transmuted into Base64 or Tensor formats and prepared for injection into the Agent's context.
  4. Tokenization: By optimizing the image before it reaches the Animator, the Prism significantly reduces the token cost and memory pressure of visual reasoning.

IV. The Vision Toolset (A Granted Power)

Vision is not a passive stream; it is a temporary power granted to Agents.

  • The Granting: When the vision.coven is active, the Dispatcher endows the Agent with a specialized toolset.
  • The Arsenal: This may include analyze_image(path, query) for general reasoning or extract_text_from_image(path) which routes to the dedicated OCR Rune.
  • The Revocation: When the Orchestrator banishes the Vision Coven, these tools vanish from the Agent's perception, ensuring it cannot attempt to use a capability the machine no longer possesses.

The Memory Burden

Visual tokens are heavy. Ingesting multiple high-resolution images can rapidly consume the context window. The Prism works alongside the Context Manager to prune older visual data once the "Observation" has been converted into "Textual Memory."