jef

n8n node for JEF — typed decisions with calibrated confidence. Route, score and judge without generating text.

Package Information

Released: this month
Downloads: 0 weekly / 0 monthly
Latest Version: 0.2.0-rc.4
Author: Mitake Cyber Security Dev Dept

Available Nodes

Documentation generating…

JEF icon
JEF
Typed decisions with calibrated confidence. Route, score and judge without generating text.

Documentation

n8n-nodes-jef

An n8n node for JEF — typed decisions with calibrated
confidence. Route, score and judge without generating text to parse.

npm install n8n-nodes-jef

Or install it from Settings → Community Nodes as n8n-nodes-jef.

Credentials

One credential, JEF API: the base URL of your server, and an optional bearer
token for when JEF sits behind an authenticating proxy. The credential test hits
/healthz, so a wrong URL fails when you save it rather than when a workflow
runs at 3am.

Operations

Operation Use it for
Run Scene The usual case. Runs a whole playbook and returns its decision trace.
Ask Choice One routing or classification decision.
Ask Score One rating against levels you define.
Ask Yes/No One boolean judgement.
Ask Many Several questions in one call.
Health Gate the workflow on whether this server can be trusted.

Ask everything at once

Every question in a request is evaluated in parallel against the same state, and
the state is read only once. Five questions cost barely more than one, so Ask
Many followed by an IF/Switch beats one HTTP round trip per branch:

{
  "false_positive": { "type": "noul", "instructions": "這則告警是否為已知的良性樣態？" },
  "urgent":         { "type": "noul", "instructions": "是否需要立即處理？" },
  "team": {
    "type": "choice",
    "instructions": "應由哪一個團隊處理？",
    "criteria": { "soc": "一般資安監控事件", "appsec": "應用程式漏洞", "infra": "基礎設施" }
  }
}

Then branch on {{ $json.answers.team.choice }} or
{{ $json.answers.urgent.noul }}.

Run Scene output

{{ $json.action }}          // 'assign' | 'escalate_human' | 'close' | ...
{{ $json.human_review }}    // whether a person must see this
{{ $json.state_encodes }}   // 1, for the whole playbook
{{ $json.layers_skipped }}  // never reached — not the same as inconclusive

Before you automate on it

confidence is how peaked the distribution is, not the chance of being right.
A model can be decisive and wrong. Scene gates read p_correct — the calibrated
mapping of confidence onto observed correctness — and refuse to fire when the
server cannot supply it, which routes the case to a human instead.

So check Health first. calibrated: false means no threshold on that server
means what it looks like, and test_backbone: true means the answers are
well-formed and carry no semantics at all.

License

Apache-2.0

Discussion