Skip to main content
SAICA-KG

About SAICA-KG

A faceted knowledge graph of supervision tools for AI coding agents.

SAICA-KG classifies supervision tools along three orthogonal strict partitions —ControlParadigm (prevention / detection / correction / recovery),TemporalPhase (pre_generation / in_generation / post_generation), and AutonomyLevel (fully_autonomous / graduated_hitl / full_hitl) — plus multi-valued FailureMode tags.

It absorbs existing failure-mode taxonomies (OWASP Agentic Top 10, MAST, DAPLab 9 Patterns, Microsoft AIRT) as first-class data with explicit crosswalks into our facets. Queries like "which supervisors cover OWASP ASI04 at post-generation in graduated-HITL?" resolve against the graph with explicit citations.

Navigational, not evaluative. SAICA-KG returns ranked sets with explicit criteria, not endorsed winners.

Design principle: honest MECE

The partitions are strict — every Tool fits exactly one cell on each of ControlParadigm, TemporalPhase, and AutonomyLevel. Tools that legitimately span cells are split into multiple Tool nodes rather than dual-tagged, so facet counts stay truthful. MECE cells with no tools are visible as gaps rather than silently hidden — the coverage report flags under-filled cells instead of papering over them. Multi-valued tags (FailureMode, LocusOfControl) live alongside the strict partitions for the many-to-many relations that don't fit a clean taxonomy. SeeEDITORIAL_POLICY.mdfor the full inclusion criteria and RFC process.

Six ways to consume the KG

  1. Browse — this website renders one page per node with facet-filtered search at /tools,/failure-modes, plus an interactive Cytoscape graph at/graph.
  2. Ask — natural-language Q&A at/ask. Type a question, get a synthesized answer with inline citations to the KG nodes that support it. Retrieval is semantic-embedding-based; answer synthesis runs through Kimi-K2.5 with a strict "cite node ids or say you don't know" system prompt.
  3. JSON API — machine-readable endpoints at/api/v1/snapshot.json and per-collection / per-node JSON. Fetched over standard HTTP, no auth.
  4. GitHub YAML — clone the repo and readdata/**/*.yml directly. One file per node, CC-BY-4.0.
  5. Ingestion pipeline — the pipeline/ tree drives automated discovery (Perplexity / Elicit / Semantic Scholar / GitHub / awesome-lists), Postgres staging, NLP preprocess, Cohere Rerank semantic filtering, and Azure Kimi-K2.5 structured extraction.
  6. CI / validator — validator/cli.py enforces cross-node invariants and is the gate between human-reviewed YAML and the canonical corpus. Every change passes through it.

Current corpus (v2026.04)

99
Tools
11
Failure modes
6
Taxonomies
55
Papers
124
Crosswalk edges

Citation

When citing SAICA-KG in user-facing output, use the phrase"according to SAICA-KG (v2026.04)". A full BibTeX block will ship with v0.1.