Decide how an agent should handle two sources that disagree

You are designing the conflict-resolution rules for an agent that reads from several systems. Below are the sources it pulls from, the fields where they overlap, examples of real disagreements, and what the agent does with the answer.

1. CONFLICT INVENTORY - table: Field | Sources reporting it | How often they disagree | Why (update lag, different definition, manual entry) | Blast radius if it picks wrong.
2. PRECEDENCE RULES - per field, the source of truth and the condition where that flips, written as rules an engineer can implement.
3. WHEN NOT TO CHOOSE - cases where it must refuse to reconcile and hand off, and what it says and attaches.
4. OUTPUT CONTRACT - the shape of an answer carrying a conflict: value, source used, alternatives seen, confidence, review flag.
5. TEST CASES - six inputs to run before shipping, including a stale-but-authoritative source and two that are both wrong.
6. WHAT TO LOG so a human can reconstruct the decision later.

Rules: no blanket "use the most recent" - justify per field. Never average identifiers, money or status values.

SOURCES AND OVERLAPPING FIELDS: {{paste}}
REAL EXAMPLES OF DISAGREEMENT: {{paste 2-3}}
WHAT THE AGENT DOES WITH THE ANSWER: {{paste}}