Granite Guardian
Foundation-model content moderation: jailbreak, harm, bias, plus RAG groundedness, relevance, and answer-relevance checks. Runs via Ollama. Off-the-shelf safety component reused across agents.
Source ↗General safety + science-specific guardrails
Composable components teams declare at design time, run during build, and keep active at runtime. Generic LLM safety on one side, science-specific reasoning on the other. Both are separate components in the AKD stack.
What guardrails detect
A non-exhaustive view of failure modes guardrails actively check for in scientific agent outputs:
- Hallucination
- Attribution gaps
- Unsupported claims
- Outdated confidence
- Overgeneralization
- Lack of grounding
- Multidisciplinary reasoning failures
- Consistency issues
General safety
Off-the-shelf LLM safety components reused across every agent in the catalog.
Defense-in-Depth
Layered review pattern: input filters, intermediate-step checks, and output verification. A composition pattern, not a single component; AKD applies it across agent boundaries rather than gating only at one point.
Source ↗Science-specific guardrails
Components built specifically for scientific agent outputs: claim-level factuality, NASA risk taxonomies, and domain-aware compliance reasoning.
FactReasoner
Claim-level factuality reasoning. Decomposes agent outputs into claims, scores attribution and supporting evidence, and surfaces unsupported assertions before they reach the user.
Source ↗Risk Agent
LLM-judge that classifies and explains risks in agent outputs against two taxonomies: the IBM Risk Atlas and the NASA Science Literature Risk taxonomy. Domain-aware, importance-weighted, DAG-based evaluation graph.
Source ↗