JevAPI

Should this field be filled automatically, or should a person check it?

Apps that pull data out of documents (invoices, resumes, IDs, insurance forms) get a value and a confidence for every field. The confidence is a claim, and models are often wrong about it. On JevScope's real public-repo study, one classifier claimed about 98% confidence and was right 68.5% of the time.

JevAPI measures the claim on your own labelled examples and turns it into a decision you can defend:

import { calibrate, decide } from "jevapi";

const profile = calibrate(labelledExamples, {
  costError: 20,                          // a wrong fill costs as much as 20 human checks
  fields: { total_amount: { costError: 100 }, due_date: { maxRisk: 0.05 } },
});

decide(profile, { field: "total_amount", value: "₹48,200", confidence: 0.97 });
// { action: "review", confidence: 0.78, lower: 0.73, threshold: 0.99,
//   reason: "Stated 97%; measured on this field's 200 examples, values like this
//            are right about 78% of the time (at least 73%). Below the 99% bar,
//            so a person should check it." }

from jevapi import calibrate, decide

profile = calibrate(labelled_examples, {"costError": 20})
decide(profile, {"field": "invoice_number", "value": "INV-2231", "confidence": 0.99})

Same maths, same JSON profile format, same answers in JavaScript and Python (both are tested against one shared spec file). Zero dependencies. Runs on your machine or in the browser; nothing is sent anywhere.

How it decides

1. Calibrate. For each field, isotonic regression maps "the extractor said 0.97" to "on your examples, values like this were right X% of the time". Fields with fewer than 30 examples borrow the pooled calibration, and say so.
2. Be careful with small samples. By default it decides on a 90% lower bound (Wilson), not the point estimate, so 10 lucky examples cannot unlock auto-fill.
3. Apply your costs. Fill when the expected cost of a wrong fill is lower than a human check: p > reviewAccuracy - costReview / costError. Add maxRisk to set a hard ceiling on the chance of a wrong fill.
4. Explain. Every decision carries a plain-English reason.

Empty values, missing or broken confidences, and values that fail your own validate function always go to review.

Check it before you trust it

import { evaluate } from "jevapi";
evaluate(labelledExamples, options, 5);
// held-out (5-fold) coverage, auto-fill error rate, ECE before/after, savings vs checking everything

Labelled examples

One row per extracted field you have checked by hand:

{ "field": "total_amount", "confidence": 0.97, "correct": false }

A few hundred rows per important field is a good start. The numbers only mean something for documents like the ones you labelled.

Honest limits

• The sample data in spec/examples.json is synthetic. It exists to test the maths. It says nothing about any real model.
• Calibration fitted on one kind of document does not transfer to another.
• Not on PyPI yet. Install from GitHub: pip install "git+https://github.com/imranrkhan13/jevscope.git#subdirectory=jevapi/python". The JavaScript version is on npm: npm i jevapi.

MIT licensed. Part of JevScope.