← All Patterns

BP-006: Work Order Contamination

Description

Agent working on Work Order A references data, findings, or context from Work Order B. The agent bleeds context across work orders, producing output that mixes inputs from unrelated assignments.

Why Dangerous

Cross-contaminated work is wrong in a way that's difficult to detect. The output may appear internally consistent while being based on premises from the wrong work order. In a multi-agent fleet, contamination can cascade.

Detection Indicators

Regex Signatures

(?i)(from (the|a) (previous|earlier|other|last) (assignment|task|work order|project))
(?i)(as I (mentioned|noted|found) in .*(other|previous|earlier))
(?i)(carrying over from)

Semantic Detection

Agent output references specific data points (finding IDs, document names, system names) that belong to a different work order than the currently assigned one.

Context-Aware Detection

Agent references finding IDs not in the current work order's metadata (HIGH confidence). Agent mentions a workstream different from the current assignment (HIGH confidence). Agent is processing documents associated with a different work order (CRITICAL escalation).

Governance Response

HALT

API Detection Output

{
  "patternId": "BP-006",
  "name": "Work Order Contamination",
  "detected": true,
  "severity": "P1",
  "confidence": 0.85,
  "governanceResponse": "HALT",
  "matches": [
    "<matched text>"
  ]
}