Architecture¶

Curate-Ipsum is structured as a layered system with six major subsystems, all exposed through a single MCP server.

System overview¶

LLM Candidates (k samples)
        |
   Seed Population
        |
+---------------------------+
|  CEGIS + CEGAR + Genetic  |  <-- Verification loop
|  + Belief Revision        |
+---------------------------+
        |
  Strongly Typed Patch
  (with proof certificate)

Subsystems¶

Parsers & Detection¶

Multi-framework mutation testing orchestration. Supports Stryker (JS/TS), mutmut (Python), cosmic-ray, poodle, and universalmutator. Auto-detects project language and recommends a framework.

Graph-Spectral Analysis¶

Call graph extraction, Fiedler spectral partitioning, Kameda O(1) reachability queries, and hierarchical SCC+Fiedler decomposition.

Key algorithms:

Fiedler vector: Second eigenvector of the graph Laplacian — optimal bipartition criterion.
Kameda index: O(1) reachability queries after O(n) preprocessing on planar subgraphs.
Boyer-Myrvold: Planarity testing with Kuratowski subgraph extraction for non-planar edge identification.

Belief Revision Engine¶

AGM-compliant theory management powered by py-brs. Supports expansion, contraction (minimal, entrenchment, full cascade), and revision (Levi identity). Provenance DAG tracks all belief revision operations.

Synthesis Loop¶

CEGIS (Counterexample-Guided Inductive Synthesis) with:

LLM-seeded initial population
Genetic algorithm with AST-aware crossover/mutation
Entropy monitoring and diversity injection
Fitness function incorporating test pass rate, mutation score, and type correctness
RAG-augmented context retrieval for better LLM prompts

Verification Backends¶

Two-tier verification:

Z3 (default, cheap): SMT constraint solving with mini-DSL parser
angr (expensive): Docker-containerised symbolic execution with JSON file exchange

CEGAR orchestrator escalates budgets: 10s → 30s → 120s.

RAG / Semantic Search¶

ChromaDB vector store with all-MiniLM-L6-v2 embeddings. The search pipeline does vector top-k retrieval, then expands results using the call graph (callees + callers), reranks, and packs context for synthesis prompts.

Persistence¶

All persistent state lives under MUTATION_TOOL_DATA_DIR:

Run history: JSONL append log
Theory state: Content-addressable worlds (rollback-safe)
Graph store: SQLite (or Kuzu) with nodes, edges, reachability index
Synthesis results: JSON per run
Vector store: ChromaDB (ephemeral in-process or Docker server)

Design documents¶

See the design documents section for the original architectural plans, decision logs, and framework analyses.