How Decisions Work

You do not need deep Jev knowledge to use this connector. Jev turns JSON state and typed questions into numbers and labels. The connector and your flow own the if: thresholds, routes, and side effects.

Two layers

Layer	What it does	Runs where
Decide / Select	Call the model and return typed answers (noul, choice, score, …).	Provider (or mock)
Policy / Filter bands	Turn those answers into keep/drop/route actions your flow can branch on.	Inside Mule — no extra HTTP call

A typical governance flow is: Evaluate (or a shortcut) → Apply Policy → Mule Choice on action / routeKey. Filtering many items is a separate path: Filter returns kept / uncertain / dropped in one operation.

What an answer looks like

Question type	Payload shape (simplified)	How the flow usually uses it
Noul (yes/no)	noul = probability of “yes” (0–1)	Compare to a threshold, or three bands (yes / no / middle)
Choice	choice, option probabilities, optional confidence	Route on the label; demand higher certainty for riskier options
Score	score, level, level distribution	Accept or review named levels

Evaluate returns named answers under payload.answers.<id>. Shortcuts (ask-noul, choose, score) return a single answer shaped as question id result. A question-set policy keyed by real ids will not see that shortcut answer — see fails closed below.

Policy: ACCEPT, REVIEW, REJECT

Apply Policy reads answers plus local thresholds and returns:

{ "action": "ACCEPT|REVIEW|REJECT", "routeKey": "…", "reasons": [], "perQuestion": {} }

Across questions, the most cautious action wins: REJECT > REVIEW > ACCEPT.

Prefer three-band Noul

For yes/no questions, prefer:

"urgent": {
  "yesAbove": 0.7,
  "noBelow": 0.3,
  "onYes": "ACCEPT",
  "onNo": "ACCEPT",
  "onUncertain": "REVIEW"
}

A clear “no” is usually still fine for the rest of the decision (onNo: ACCEPT). Legacy rejectBelow turns a clear “no” into REJECT for the whole decision — Validate Question Set warns about that.

Per-option Choice

Riskier Choice options can demand more certainty:

"team": {
  "minProbability": 0.7,
  "onNoMatch": "REVIEW",
  "options": {
    "other": { "action": "REVIEW", "minConfidence": 0.7 }
  }
}

Stable routeKey

Set policy.routeQuestion to the Choice id that should supply routeKey (for example "routeQuestion": "team"). If you omit it, the connector uses the first Choice answer in the decision — reordering questions can change routing.

Policy fails closed

Since 1.0.2, Apply Policy returns REVIEW, never a silent ACCEPT, when it has nothing trustworthy to judge:

Situation	Expect
Empty or wrong decision payload	REVIEW — decision has no answers to judge
Policy rule whose question has no answer (common: shortcut result vs question-set ids)	REVIEW — <id>: no answer to judge
Rule keys that do not fit the answer type	REVIEW — rule does not fit a '…' answer

Invalid policies raise TYPESAFE:INVALID_QUESTION_SET before evaluation (misspelt keys, wrong question type, inverted bands, and so on). Run Validate Question Set on the file first — it checks the same policy block.

Raise on review / reject

When Raise on review or Raise on reject is enabled, Apply Policy raises typed TYPESAFE:BELOW_THRESHOLD or TYPESAFE:REJECTED. Error handlers for those types match. Before 1.0.2 those raises were rewritten to MULE:UNKNOWN.

Filter: kept, uncertain, dropped

Filter asks one Noul per item and partitions the array:

Band	Rule
kept	noul >= threshold
dropped	noul < dropBelow
uncertain	in between, when dropBelow < threshold

If you leave Drop below equal to Threshold (the default), there is no middle band — every item is kept or dropped.

Items are sent in state.items and each question refers to items[i]. Item text is not pasted into the instructions (TypeSafe packing). Attributes succeeded means kept count on this operation.

Drift on Noul-only traffic

On Drift Detected watches no-match rate, mean confidence, and answer distribution. Noul answers have no provider confidence field, so the connector uses certainty |noul − 0.5| × 2 and buckets answers as yes / no / uncertain. Noul-only question sets still drive drift the same way Choice/Score do.

What not to expect

• Policy and validation do not call the provider.
• Apply Policy does not invent answers — wire Evaluate (or the matching shortcut shape) first.
• Filter does not apply a question-set policy block; use Apply Policy after Evaluate for full governance.
• Official Jev docs describe the model and API; this connector’s thresholds, fail-closed rules, and band behavior are documented here and on the operation pages.

See also

• Apply Policy — actions, fail-closed, validation, routeQuestion
• Filter — packing and three bands
• Validate Question Set — questions and policy checks
• Sources — drift, budget, failover
• Set Up — install and first question set
• Jev introduction — optional deeper reading on System One