l4-jev-workflows
Loopfour Studio features rebuilt with Jev: code finds candidates, Jev picks with confidence, code assembles verbatim results.
Overview
l4-jev-workflows rebuilds features of the Loopfour Studio workflow product using Jev instead of an LLM. The pattern throughout is that code finds candidate answers, such as every amount, date, term or label a document could mean; Jev picks one typed answer per question with a probability, sending low-confidence answers to a person; and code assembles the result, copying each value verbatim with its source line for an audit trail. Because Jev never writes text, a total, date or block name cannot be invented.
It offers a workflow builder that turns a plain-language description into a Loopfour workflow by having Jev file each part under a block from the workspace's live catalog, so non-existent blocks are never used; document extraction against a JSON Schema; classification with a review flag; and a 'race' that sends the same input to Jev, Claude Opus and OpenAI's Decisions API and times them live. An optional cascade sends only uncertain answers to an LLM and keeps an answer only if it quotes the document and Jev confirms it. A live site is available.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which candidate value (amount, date, term) is the correct answer for this field? | — |
| Choice | Which label(s) fit this document? | — |
| Choice | Which catalog block does this part of the description map to? | — |
How it works
- Code extracts candidate values or blocks
- Jev picks one per question with a probability
- Low-confidence answers go to review or an optional LLM cascade
- Code copies values verbatim with source lines
- Workflows are created through the Loopfour API
Strengths
- Values cannot be invented
- Workflow builder uses only real catalog blocks
- Live race compares with LLMs
- Audit trail per value
Caveats
- Specific to Loopfour Studio
- Candidate extraction quality limits accuracy
- Requires a TypeSafe key
Getting started
Try the live site at https://l4-jev-workflow-builder.vercel.app; see workflow-builder/README.md for setup.At a glance
- Type
- Project
- Status
- prototype
- Language
- TypeScript
- Runtime
- Next.js app on Vercel
- Has code
- yes
Use cases and integrations
How curated lists describe it
Loopfour workflow steps rebuilt with Jev (TypeSafe) instead of an LLM.
wh000wh000/awesome-jev-live
Frequently asked questions
Can Jev invent a value?
No. It only picks among candidates found in the document; code copies the value verbatim.
What happens when Jev is unsure?
The answer goes to a person or, optionally, to an LLM whose answer must quote the document and be confirmed.
What is the race?
A live timing comparison of Jev, Claude Opus and OpenAI's Decisions API on the same input.
Related projects
macOS bookkeeping clerk booking supplier invoices: Jev picks each click from a closed list; a deep model rewrites the playbook.
ServiceprototypePythonClassifies tax PDF pages into 261 IRS forms and 7 page kinds with one Jev request per page and a confidence gate.
ProjectproductionTypeScriptRust tool that asks Jev which PDF pages genuinely need OCR, extracting text pages locally and sending only the rest to OCR.
ProjectbetaRustLocal library that turns an image into a transparent text state (OCR, boxes, labels) so Jev can decide over screenshots.
ProjectbetaPythonRust streaming event bus whose routing is decided by Jev, one question per subscription, with deliver, review or drop thresholds.
ProjectbetaRustWindows office workbench with search, OCR and batch renaming, where Jev optionally helps pick extracted contract fields.
ProjectbetaRust