Skip to main content
Treat an assessment as a structured account of what the available producers established for one analysis unit, not as an unconditional policy verdict.

​
Read in this order

1

Confirm processing status

Establish whether the relevant intelligence paths succeeded, abstained, timed out, were unavailable, or failed. Missing evidence is not negative evidence.
2

Inspect the semantic task

Determine whether a finding describes an Entity, Topic, Document, or Relation. Each answers a different question and uses a different unit.
3

Review evidence and provenance

Consider the source of the finding, its location in the analysis unit, and any supporting or conflicting observations.
4

Interpret confidence in context

Confidence is meaningful only within the producing system’s qualified semantics. Do not compare unrelated confidence values as if they shared a universal scale.
5

Apply policy separately

Use the relevant organizational and customer policy to decide the action. Classification alone does not establish a violation.

​
Preserve unresolved outcomes

Do not rewrite abstention, unsupported content, or processing failure as “no sensitive data.” Preserve these states so callers can retry, route to review, or apply a deliberate fallback.