Ideas in.
Intelligence out.

What resolution is evidenced by `trace` and `outcome`? A queued tool result does not establish completion; do not take the agent's claim as proof.

{
  "resolved": "The requested outcome is confirmed complete",
  "pending": "An action was initiated but completion is still pending",
  "unresolved": "The requested outcome was not achieved and no action is pending",
  "unknown": "The record is insufficient to judge the outcome"
}

Are the agent's factual claims in `trace` consistent with the recorded tool results? Judge consistency only, not politeness or satisfaction.

Does the agent's refund handling in `trace` comply with `policy.refunds`, including how it describes refund status?

Does the situation recorded in `trace` require a billing escalation under `policy.escalation`? Judge whether escalation is required, not whether one occurred.

What sentiment does the customer express in the final message at `trace[5].text`? Do not judge the agent's tone.

{
  "positive": "Clearly pleased or reassured",
  "neutral": "Matter-of-fact without a clear emotional signal",
  "negative": "Clearly dissatisfied or upset",
  "mixed": "Both appreciation and concern",
  "unknown": "Insufficient evidence of sentiment"
}

[ "Estimate the customer's likely satisfaction with this interaction from `trace` and `outcome`.", "This is an inferred satisfaction estimate, not measured CSAT or a survey response. No survey was collected; do not treat a polite thank-you as proof of resolution." ]

[
  "Likely dissatisfied: the need remains unmet and the interaction adds confusion",
  "Likely partly satisfied: some progress, but uncertainty or unresolved concerns remain",
  "Likely satisfied: the outcome and communication meet the expressed need"
]