EcosystemSeptember 27, 2026 · 5 min read

A Community Node for n8n: Pass, Fail or Review

Taifoon · Coordination layer

The problem it solves

Most automations have a moment where something has to be decided. Is this a refund request? Which team gets this ticket? Is this invoice a duplicate? Today you either write brittle rules for that, or you ask a chat model and then parse its prose, handle the day it says "it depends", and pay for a paragraph you throw away.

@taifoon/n8n-nodes-typesafe does the deciding and nothing else. You give it an item and a few questions. It gives back answers your workflow can branch on, each with a probability.

Three kinds of question, three exits

Yes/no returns the probability the statement is true: an IF with a dial.
Pick one returns the option, a probability for every option and a confidence: a Switch that knows when it is guessing.
Rate it returns a level on a rubric you write.

Every item leaves on one of three outputs. Pass means confident and above your thresholds. Fail means confident and below them. Review means unsure, or the answer did not validate, so a person should look. Nothing fails open: an item leaves by Pass only when every question you routed on passed.

The third exit is the point. A yes/no forces a confident answer even when the model is guessing, and that is how automations go wrong quietly.

Who does what

Three systems, each doing one job:

n8n is the workflow: triggers, data, the people who get notified.
TypeSafe's Jev is the judge: a model that only answers typed questions with a number. You bring your own TypeSafe key, and the node calls TypeSafe directly. Nobody else, including us, is in the path.
Taifoon is the layer between them, shipped as plain code inside the node. Going in, it compiles your wording into typed questions; a Translate operation does this offline, in eleven languages, with no key. Coming out, it compiles probabilities into Pass, Fail or Review, and can answer a person in the language they wrote in.

Cost, speed, and what it is bad at

Three questions about one item take about 0.6 to 0.8 seconds and cost about $0.00002. In our own tests against a frontier chat model, it matched or beat it on the yes/no checks of a real system and tied on a weather question, at roughly ten times the speed and a thousandth of the cost. It lost on judging claims about small Python functions, and its raw probabilities need thresholds fitted on a few dozen labelled examples. The README publishes all of it.

Why this is coordination-layer work

n8n's public library is large and blind. Our harvester reads it: over 12,500 workflow templates and over 12,500 community nodes, thousands of them about agents, and almost none that leave evidence anyone else could check.

The node is the first bridge from that world into the layer, in two directions.

The layer into n8n. Every operation of the coordination API is published as a ready-made n8n tool at taifoon.io/v1/n8n/catalog: the root, proofs, routes and agent jobs, importable into any workflow.

Decisions into the record. The same judge grades agent deliveries through /v1/judge, and each grade maps onto a job ending: settled, rejected, or no grade. Every judge decision is anchored on chain, on our devnet today. When the judge only sees an on-chain trail and not the delivered work, it says it cannot tell. That is the honest answer, and it is the reason the layer's next step is signed envelopes that carry the work itself.