Jev Case Study: Invoice Approval Gate (Fallback Code)

TL;DR: This Jev example builds an invoice approval gate on one judgment question: Jev 1.13 answers “approve under policy?” with yes/no/unclear plus a confidence value, and the gate auto-approves only when the answer is yes with confidence at 0.8 or higher. Everything else falls back to a human. Full Python below.

Scenario

Our accounts payable team processes a few thousand invoices a month. Most are unremarkable — under $500, matching a purchase order, correct vendor. A junior accountant still opened each one, checked three fields against the PO, and clicked approve: ten seconds of skilled attention per invoice, thousands of times a month.

The policy is simple enough to state in one sentence, so we turned it into one judgment question:

Should this invoice be auto-approved under policy (amount under $500, matching a purchase order)?

And one fallback rule, which is the actual point of this case study:

• answer: "yes" and confidence >= 0.8 → auto-approve, archive the rationale.
• answer: "no" → reject and notify AP with the rationale.
• answer: "unclear" or confidence < 0.8 → human review queue.

The gate never auto-approves on a whisper. A judgment model that says “yes, 0.55” is telling you the invoice sits in a gray zone — a missing PO reference, an unusual line item, a vendor name that almost matches. Those are exactly the invoices a person should see, and exactly why Jev — TypeSafe AI’s System One judgment model (September 2026) — returns confidence as a first-class field in its typed output rather than burying it in prose. Calls go through the OpenRouter OpenAI-compatible endpoint.

Request

{
  "model": "typesafe/jev-1.13",
  "messages": [
    {
      "role": "system",
      "content": "Judgment question. Should this invoice be auto-approved under policy (amount under $500, matching a purchase order)? Answer yes, no, or unclear. Reply with JSON only: {\"answer\": <yes|no|unclear>, \"confidence\": <0-1>, \"rationale\": <one sentence>}"
    },
    {
      "role": "user",
      "content": "Invoice INV-2041 from Northwind Office Supply\nAmount: $412.80 | PO: PO-8877 | Terms: Net 30\nLine items: 12 ergonomic chair mats, matching PO-8877 pricing."
    }
  ]
}

Response

Example fixture — illustrative output, not a live capture:

{
  "answer": "yes",
  "confidence": 0.97,
  "rationale": "Amount is under $500, the invoice references PO-8877, and line items match the purchase order pricing."
}

The fixture is the boring case on purpose: clean yes, high confidence, auto-approved. The gate earns its keep on the cases where either field degrades — see the fallback table below.

Reproduce

Save as gate.py, set OPENROUTER_API_KEY, and run python3 gate.py:

import json, os, urllib.request

URL = "https://openrouter.ai/api/v1/chat/completions"
MODEL = "typesafe/jev-1.13"  # confirm exact slug on the OpenRouter model page
HUMAN_REVIEW_CONFIDENCE = 0.8

SYSTEM = ('Judgment question. Should this invoice be auto-approved under '
          'policy (amount under $500, matching a purchase order)? '
          'Answer yes, no, or unclear. Reply with JSON only: '
          '{"answer": <yes|no|unclear>, "confidence": <0-1>, "rationale": <one sentence>}')

def gate(invoice_text):
    body = {"model": MODEL, "messages": [
        {"role": "system", "content": SYSTEM},
        {"role": "user", "content": invoice_text}]}
    req = urllib.request.Request(URL, data=json.dumps(body).encode(),
        headers={"Authorization": "Bearer " + os.environ["OPENROUTER_API_KEY"],
                 "Content-Type": "application/json"})
    with urllib.request.urlopen(req) as res:
        payload = json.load(res)
    result = json.loads(payload["choices"][0]["message"]["content"])
    if result["answer"] == "yes" and result["confidence"] >= HUMAN_REVIEW_CONFIDENCE:
        return "auto-approved", result
    return "human-review", result

invoice = """Invoice INV-2041 from Northwind Office Supply
Amount: $412.80 | PO: PO-8877 | Terms: Net 30
Line items: 12 ergonomic chair mats, matching PO-8877 pricing."""

print(gate(invoice))

Key parameters

Notes

• Responses are example fixtures; verify field names against official docs before relying on them. Official API details live at typesafe.ai — see the System One announcement post.
• Confirm the exact model slug (typesafe/jev-1.13) on the OpenRouter model page before shipping.
• Parse defensively in production: if confidence is missing or not a number, treat it as human-review, never as approval.
• Log the confidence distribution weekly. A sudden shift toward low confidence usually means your invoice mix changed, not the model.
• Keep the policy sentence and your written AP policy in sync — the model only knows the rule you put in the system prompt.
• Applies to Jev 1.13 (released September 2026) via the OpenRouter demonstration channel.