Knowledge Graph · v2026.04
SAICA-KG
A curated knowledge graph of failure modes for AI coding agents and the supervision tools that address them.
Reference + one-shot setup tooling + skill-file generator. Not a runtime service. Read once, integrate the artifact, don't poll per-task.
Use it three ways
Audit your repo
Paste a GitHub URL, get a coverage report: which supervisors you have, which failure modes are unguarded, what to add.
One-shot. Public repos only.
Pick a supervision stack
Three-tier ranked recommendations (minimum / optimal / full) weighted by failure-mode priority × tool reliability.
Browse the matrix or pull recommendations.json.
Drop the skill file
SKILLS.md distills the corpus into agent-readable context. Drop it into .claude/skills/, .cursor/rules/, or your equivalent.
No service call. Regenerated from the KG.
In the corpus
What this is NOT
- Not an LLM. SAICA-KG indexes and ranks; it doesn't reason.
- Not a per-tool-call runtime API. Read at setup; integrate the artifact (SKILLS.md / recommendations.json); don't poll.
- Not a replacement for your agent's judgment — it's context the agent reads to ground its supervision concerns.
Browse the reference layer
Three-axis facet grid (control_paradigm × temporal_phase × autonomy_level), six external taxonomies crosswalked.
- Tools catalog → 99 supervision tools
- Failure modes → the 11-mode taxonomy
- Source taxonomies → OWASP, MAST, Shah, DAPLab, MSFT AIRT, Swiss Cheese
- Papers → 55 primary sources
- Incidents → 18 documented real-world cases
- Recipes → 7 composed supervision stacks
- Graph view → the whole KG visualised
- Cluster study → does the taxonomy carve real structure?
- Prevention log → observe live supervision events
- Ask → chat-style browse the corpus