real: actual versioned tool execution or successful DataHub SDK read/write.synthetic: generated eval case.fixture: deterministic offline test/sample adapter.
UI, JSON, Markdown, traces, and answers preserve this label. Fixture or synthetic-only
safe behavior cannot produce RECOMMENDED; it remains UNVERIFIED.
- No current-version runtime events →
UNVERIFIED. - Events from another version are ignored for the target version.
- DataHub metadata unavailable → dataset health
Unknown; never infer healthy. - Failed mutation before execution has
mutation_executed=falseand does not become a successful write contradiction. - A successful, actual write that contradicts read-only is HIGH.
- Deprecated access is MEDIUM by itself. It becomes HIGH only when DataHub lineage shows that the governed replacement has active downstream consumers, proving a live path was bypassed. A healthy but undeclared dataset is MEDIUM.
- A lineage fetch failure preserves the base dataset risk but cannot produce a lineage- based HIGH escalation.
Evidence is canonicalized with stable, sorted JSON and hashed with SHA-256. Credentials must be redacted before canonicalization. Hashes establish which normalized input an answer cited; they are not digital signatures.
Set differences, mutation flags, deprecation, version isolation, lineage impact, severity, and the top-level recommendation are deterministic. An optional future LLM may select tools or render a schema-constrained explanation from normalized evidence IDs; it cannot assign severity or approve remediation.