OptionalbatchShared across every record from the same top-level call (e.g. all chunks in one scanRAGChunks call). Undefined for single-content calls.
Unique categories among detected threats.
Agent Armor's classification of this scan — see the interface doc comment above.
Scan duration in ms.
OptionalexceptionPopulated only when decision === 'exception' — both fields required (cannot omit reason or actor).
OptionalindexChunk index (scanRAGChunks) or turn index (scanSession). Undefined for single-content calls.
OptionalmlML model version, present only when the ML classifier ran for this scan.
Pattern database version that produced this decision.
Unique id for this specific decision (one per chunk/turn, not per API call).
Record format version, for forward compatibility.
Which SDK entry point produced this record.
Confidence-threshold preset active for this scan.
Per-threat summaries — see AuditThreatSummary.
ISO 8601 timestamp of when this scan decision completed.
A durable, structured record of one scan decision — the substrate for SOC2/ISO27001-style audit trails (#38).
decisionis Agent Armor's own classification derived fromriskLevel, NOT a guarantee of what your application actually did with theScanResult— that decision happens in your code, after the scan call returns, where the SDK can't observe it.