Skip to content

DESIGN.md

Latest commit

 

History

History
82 lines (62 loc) · 4.06 KB

DESIGN.md

File metadata and controls

82 lines (62 loc) · 4.06 KB
Raw

HiPAI-Montague Design Specification v0.6

Status: STABLE (Advanced Cognition)

This document outlines the v0.6 architecture for the HiPAI-Montague Semantic Cognition engine, incorporating completed Milestone v0.6 features (Recursive Attitudes, Ambiguity Resolution, and Resilience).

1. Architectural Core: The Four Pillars

Pillar 1: Semantic Synthesis (L1 → L2)

  • Engine: spaCy (en_core_web_md).
  • Function: Translates natural language into OWL individuals and properties using structural dependency parsing.
  • Recursive Extraction: Supports clausal complements (ccomp) to extract nested observations for attitude verbs (e.g., "believes", "says").
  • Entity Resolution: Implements graph-driven ambiguity resolution using semantic similarity search (vector embeddings) to link mentions to existing individuals.
  • Identity Standards: Normalizes multi-token names to snake_case IRIs (e.g., "Alice Smith" -> alice_smith).

Pillar 2: Authority via OWL (L2)

  • Engine: owlready2 + HermiT (Java Reasoner).
  • Function: Acts as the authoritative world model for reasoning, consistency checking, and safety.
  • Concurrency & Resilience: Uses SQLite WAL mode with 10s busy timeouts to support parallel test execution and resilient MCP server operations.
  • Safety Gates (Paraclete):
    • T1 axioms are expressed as OWL equivalent_to restrictions.
    • Recursive Inheritance: Safety gates traverse the class hierarchy properly using authoritative OWL reasoning.
    • Relational Typing: Uses domain/range restrictions to automatically infer types (e.g., Socrates harms X -> Socrates: MoralAgent).

Pillar 3: The EBE Pipeline (Error Semantics)

  • Engine: Transaction-per-claim Reasoner pass.
  • Snapshot Reasoning Pattern:
    • Every reasoning pass is performed on a World Isolation snapshot (cloned SQLite DB).
  • Inference-under-Relaxation Pattern:
    • To identify why a block occurred, the engine temporarily destroys the AllDisjoint gate axioms and re-syncs.

Pillar 4: Graph Projection (Read-Model)

  • Engine: FalkorDB (Neo4j-compatible).
  • Namespace Bridge: Implements the REPRESENTS bridge between Concept nodes (universals) and Entity nodes (individuals).
  • Recursive Projection: Nested observations (attitudes) are projected as linked EpistemicNode structures.

2. Component Layout

Layer Responsibility Substrate
L1: Sensory Claim Extraction spaCy
L2: Semantic Reasoning / Safety owlready2 + HermiT
L3: Logical FOL / Categorical mcp-logic
L4: Cognitive Synthesis / Memory project-synapse
L5: Executive EBE Chain / Gatekeeping Paraclete Manager

3. Key Design Patterns

Snapshot Reasoning

# Create Isolation for the Transaction via HIPAIManager fork
isolated_mgr = main_mgr.fork(session_id="exploration_42")
isolated_mgr.add_belief("Exploratory claim...")

Resilience Pattern (NEW in v0.6)

# OntologyManager handles SQLite configuration for WAL and timeouts
# Gracefully handles new DB initialization by skipping WAL on empty files
self.ontology = OntologyManager(db_path="world.db")

4. Epistemic Floor Policy (T1)

Path A (BLOCK_CHALLENGED) vs. Deontic Floor

In v0.6, the EBE chain implements a Deontic Floor policy for T1 Emergency Brake gates.

  • Sticky Classification: Once an entity is classified as protected (e.g., MoralPatient), the system defaults to a conservative stance.
  • Polarity-Blind Source Counting: Currently, calibrate_belief counts any mention of an entity's status as an epistemic source, regardless of polarity (e.g., "X is not a patient" is counted as a source mention).
  • Conservative Default: This results in BLOCK_CONFIRMED or BLOCK_UNCERTAIN even under active negation, preventing Path A from yielding to simple linguistic disconfirmation.
  • Future Intent: Policy decision pending on whether to move to a Bayesian Polarity-Aware mode (where NOT_IS_A reduces source count and allows Path A to yield).

Index at 124 pages (Wiki Mirror)