Sources

The connector provides three polling sources for its own governance signals. Business events should still enter through normal Mule sources such as HTTP, Salesforce, Anypoint MQ, or Scheduler.

All three sources require a connector configuration, connection, and Mule scheduling strategy. Their output attributes are null.

On Drift Detected

Detects a change in no-match rate, mean confidence, or Choice/Score distribution.

Parameter	Default	Description
Question set ID	Empty	Limit monitoring to one question set.
Question ID	Empty	Limit distribution monitoring to one question.
Window size	500	Decisions in each comparison window.
Baseline	FIRST_WINDOW	Compare with the first or previous full window.
Max no-match rate increase	0.10	Allowed increase above baseline.
Max mean confidence drop	0.10	Allowed decrease below baseline.
Max distribution shift	0.10	Jensen–Shannon divergence threshold from 0 to 1.

Payload:

{
  "metric": "MEAN_CONFIDENCE",
  "baseline": 0.84,
  "current": 0.69,
  "windowSize": 500,
  "questionSetId": "ticket-triage",
  "questionId": "team",
  "windowEnd": 1758996000000
}

windowEnd is epoch milliseconds (not an ISO-8601 string). At least min(30, windowSize) current samples are required. The source fires once per breach and re-arms after recovery.

<typesafe:on-drift-detected
    config-ref="TypeSafe_Config"
    questionSetId="ticket-triage"
    windowSize="500"
    maxMeanConfidenceDrop="0.10">
    <scheduling-strategy>
        <fixed-frequency frequency="60" timeUnit="SECONDS" />
    </scheduling-strategy>
</typesafe:on-drift-detected>

On Budget Threshold

Fires when calls or input tokens reach a configured percentage of their rolling-window limit.

Parameter	Default	Description
Percent	80	Percentage at which to fire.
Metric	CALLS	CALLS or INPUT_TOKENS.

Payload:

{
  "metric": "CALLS",
  "used": 806,
  "limit": 1000,
  "percent": 81,
  "threshold": 80,
  "estimatedCostUsd": 0.18,
  "windowStart": 1758996000000
}

windowStart is the rolling window start as epoch milliseconds (not an ISO-8601 string). estimatedCostUsd is derived from configured price per million input tokens and token usage in the window; it may be a small decimal when only a few calls have run. Attributes are null.

The matching budget limit must be set on the global connector configuration.

<typesafe:on-budget-threshold
    config-ref="TypeSafe_Config"
    percent="80"
    metric="CALLS">
    <scheduling-strategy>
        <fixed-frequency frequency="60" timeUnit="SECONDS" />
    </scheduling-strategy>
</typesafe:on-budget-threshold>

On Provider Failover

Emits each successful fallback after the primary route could not serve a request.

Payload:

{
  "from": "typesafe",
  "to": "openrouter",
  "reason": "FAILOVER",
  "errorType": null,
  "timestamp": 1758996000000
}

timestamp is epoch milliseconds (not an ISO-8601 string). errorType is currently always null. Attributes are null.

The source polls stored failover events using a watermark. Events are recorded only after the fallback successfully answers a stats-enabled decision.

<typesafe:on-provider-failover config-ref="TypeSafe_Config">
    <scheduling-strategy>
        <fixed-frequency frequency="60" timeUnit="SECONDS" />
    </scheduling-strategy>
</typesafe:on-provider-failover>

Privacy and persistence

Budget and drift stores retain counters and distributions—not state text. Budget and statistics stores are distributed and persistent so monitoring remains consistent across a Mule cluster.