Skip to content

ARCHITECTURE.md

Latest commit

 

History

History
90 lines (66 loc) · 3.5 KB

ARCHITECTURE.md

File metadata and controls

90 lines (66 loc) · 3.5 KB
Raw

Ghost Architecture Handbook

This document describes implemented behavior and operator vocabulary for the current Ghost stack. It is not a roadmap.

Hybrid Lens

Ghost separates fast similarity from final authority.

  • The Vulkan path uses f32 lanes for bounded approximate search. It can rank, cluster, and narrow candidates quickly, including SimHash-style semantic prefiltering.
  • The CPU path keeps u32 and structured symbolic state as the truth layer. It performs branch-heavy checks, obligation accounting, support routing, and final permission decisions.
  • A high similarity score is never proof. It is a lens for finding candidates that still need symbolic validation.

The practical rule is: GPU similarity may suggest where to look; CPU symbolic logic decides whether anything is supported.

Latent Intent Inference

intent_grounding projects user text into a small deterministic f32 concept space. The current concept families are conversation, modification, retrieval, and meta-analysis.

If a request does not cross the heavy-task confidence threshold, Ghost falls back to conversational reasoning. In that mode it does not require a bind_target_artifact obligation. Code-heavy requests still require concrete artifact grounding before they can move toward support.

This is the Intent Fallback rule: low-confidence task routing talks first, and only asks for paths or bindings when the user actually expresses code-heavy work.

Authority Ranks

Ghost treats source authority as explicit metadata, not vibes.

  • Rank 0, Sovereign or Zenith: local root authority. This is the highest-trust operator/project layer and may anchor project-specific reasoning.
  • Rank 1, Verified: reviewed project or corpus material that has been promoted by an explicit human workflow.
  • Rank 2, Public or Wiki: public corpus/reference material. It can inform answers but cannot outrank local sovereign project state.
  • Rank 3, Shadow: retained but low-trust material. It may surface as a warning or contrast candidate.
  • Rank 4, Trash: rejected material. Matches should be treated as falsehood or blocked influence, not evidence.

Authority ranks constrain support. They do not replace verifier evidence.

Zen Mode

The CLI/TUI defaults to Zen Mode for ordinary conversation:

  • user text renders in white
  • Ghost text renders in blue
  • obligations, ambiguity sets, missing-info sections, debug counters, and raw authority scaffolding are hidden

Details are explicit:

  • CLI: pass --details
  • TUI launch: pass ghost tui --details
  • TUI runtime: run /details on or /details off
  • Debug mode also enables details because it is a diagnostic path

Zen Mode changes presentation only. It does not grant support, hide engine failures from raw JSON, or change the underlying proof contract.

Glossary

SimHash64: A 64-bit deterministic semantic fingerprint used for compact approximate candidate comparison. It is useful for search narrowing, not final support.

Rune Validation: The symbolic validation pass that checks retained text/code fragments as bounded meaning units. Runes are inspectable units; they are not model tokens and do not imply neural training.

Intent Fallback: The deterministic fallback from low-confidence heavy-task routing to conversation. It prevents casual prompts such as hi, test, or broad status questions from failing on artifact-binding obligations.

Hybrid Lens: The architecture pattern where f32 Vulkan similarity finds candidates and u32/structured CPU logic decides truth, rank, and support.