The reviewer's standard

The reviewer writes the standard.

Each criterion has three parts: a pass condition, a threshold measured on your own labeled lines, and feedback in the reviewer's own words. The model is never shown the feedback. Edit it here and watch the preparer's line change.

  1. C1Names a specific driver0.43
  2. C2Explains why, not just what0.84
  3. C3Separates one-time from recurring0.15
  4. C4Points to evidence0.74
  5. C5Sizes each driver0.36
  6. RESTATESays more than the row0.85
Criteria tab · reviewer

C1 Names a specific driver

Passes when
Names the specific cause of the change, identified well enough that a reviewer could look it up: which customer, vendor, contract, hire or event, and what changed with it. Fails when the cause is a category ("timing", "growth", "higher volume"), even with an amount or dates attached, or when the only name given is where the cost was booked or who billed it.
Threshold
0.43 · Calibrated on this demo team's labeled lines · held-out balanced accuracy 1.00
Editing the feedback doesn't change what Jev is asked. Editing the pass condition or threshold would make C1 uncalibrated until it's measured again.
Flux tab · preparer
FHD-US|2026-09|MoM|5000Cloud hosting · +$48,500
StatusNeeds revision
Failed criteriaC1 (likely)
FeedbackC1: Who or what, specifically? Name the customer, vendor, contract or hire so I can check it.

The reviewer's words, exactly as written. Nothing is paraphrased, and the model never sees this text.

Where Jev fits

A yes or no for each criterion.

Jev is TypeSafe's yes or no decision model. One request per line carries a question for each criterion, and each answer is a probability set against that criterion's threshold. Until a team calibrates on its own labeled lines, a flag goes to the reviewer as a prompt, not as a failure.

data_collection: deny
OpenRouter routes it only to providers it lists as not collecting prompts. On by default.
allow_fallbacks: false
Never silently routed to a backup provider.
jev-1.13-20260917
A dated snapshot. An alias refuses to start.

Through OpenRouter. A team that calls TypeSafe directly is under TypeSafe's own policy.

Sent to JevFHD-US|2026-09|MoM|5000 · C1

Account
Cloud hosting
Basis
MoM
Prior
$409,000
Current
$457,500
Change
+$48,500 · 11.9%

Commentary

Hosting rose $48.5K because of a one-off data load during the month. One-time: the load finished Sep 24. See invoice SC-0930.

Pass condition, C1

Names the specific cause of the change, identified well enough that a reviewer could look it up: which customer, vendor, contract, hire or event, and what changed with it. Fails when the cause is a category ("timing", "growth", "higher volume"), even with an amount or dates attached, or when the only name given is where the cost was booked or who billed it.

Answer0.30below 0.43, so C1 fails

  • Entity, period or line ID
  • Anyone's email
  • The criterion's name, feedback, rationale or example

How it works

One click. Three outcomes.

Immaterial lines never reach the model. Objective failures get feedback before any model call. The rest get one yes or no question per criterion.

  1. 01Materiality gatedeterministic
  2. 02Arithmetic and duplicatesdeterministic
  3. 03Jev, one question per criterionprobability vs threshold
  4. 04Passed, or the reviewer's wordsback in the sheet

FHD-US · Sep 2026 · MoMOffice supplies

Up $1.2K

Skipped: immaterial

Under both thresholds in Settings. Never sent to the model.

FHD-US · Sep 2026 · MoMCloud hosting

Up $48.5K

Needs revision

Failed criteria: C1 (likely)

Who or what, specifically? Name the customer, vendor, contract or hire so I can check it.

FHD-US · Sep 2026 · MoMCloud hosting

Up $48.5K

Passed (attempt 2)

Orion Freight ran the initial data load for its second region. Every criterion met.

What it never does

  1. It never writes the explanation.

    The preparer writes every word. The Coach checks whether those words can be reviewed.

  2. It never invents feedback.

    A failed line gets the reviewer's own words, verbatim, from the Criteria tab.

  3. It never decides whether the explanation is true.

    That stays with the reviewer, and the preparer still talks to the business owner.

  4. It is not a control.

    No sign-off, no control evidence, and it never replaces the review.

Watch it in the sheet

The same loop, in the actual sheet.

Why Jev

Code counts. Jev judges. The reviewer writes.

  1. Code

    Counts and compares.

    Jev isn't a calculator and reads dates as plain text,1 so totals, duplicates and materiality are settled in code first.

  2. Jev

    Answers yes or no.

    It returns a probability and no text.25 In OpenRouter's test, a chat model's confidence was always near 0 or near 1, even when it was wrong. Jev's used the whole range, so a team can set the score below which a line fails, using its own labeled lines.34

  3. The reviewer

    Writes every word.

    The feedback a preparer reads is the reviewer's own, verbatim. No model drafts it, so none can drift or invent.

A chat model is better at open-ended grading and at explaining itself.3 Working Note needs neither: each criterion is a pass condition written down in advance, so every check is a yes or no question, and the explanation a preparer reads is the reviewer's own. A chat assistant earns its place earlier, helping the reviewer draft the standard in onboarding.

  1. 1TypeSafe: Jev 1.13 known weaknessesdocs.typesafe.ai
  2. 2OpenRouter: What is Jev?openrouter.ai
  3. 3OpenRouter: Jev vs LLM-as-a-Judgeopenrouter.ai
  4. 4TypeSafe: Confidence: thresholds and when to escalatedocs.typesafe.ai
  5. 5OpenRouter: Jev guideopenrouter.ai

These sources describe their own tests, not Working Note. Working Note's measurements are in the repo, on Fernhill Data, a fictional company.

Technical setup

Run your own copy.

This is the path for whoever sets it up. Nothing to sign up for to try it locally: Python 3.11 and Node 22, Postgres 17 only for the database tests. A team deploys its own copy, with its own database and Jev key, and calibrates it on its own lines; the README walks through every step. MIT license, and no real company's data: the tests run on a fictional one.

Not setting it up yet? Watch the sheet workflow or try the example. Neither needs an install.

Go to GitHubDownload .zip

git clone https://github.com/alexxxram25/working-note.git
cd working-note
make venv # one Python environment
make test # everything that needs no database