# AI use cases for Open Jev (kyegomez/open-jev)

Repo: [kyegomez/open-jev](https://github.com/kyegomez/open-jev) | Use cases: 3 | Page: https://dedupedai.com/use-cases/kyegomez-open-jev/

Open Jev is an untrained PyTorch research reconstruction of the ideas behind TypeSafe AI's Jev, which answers typed questions about nested state without generating text.

This file holds every use case of this repo. Each one starts with a "## " heading and ends with its URL.

- Run the bundled demo to see typed readouts in action: https://dedupedai.com/use-cases/kyegomez-open-jev/run-typed-readout-demo/
- Prototype typed yes/no, choice, and score questions over JSON state: https://dedupedai.com/use-cases/kyegomez-open-jev/prototype-typed-questions-over-json-state/
- Test soft-target training with RLCDLoss on your own labels: https://dedupedai.com/use-cases/kyegomez-open-jev/experiment-with-rlcd-loss-targets/

## How to run the bundled demo to see typed readouts in action with Open Jev

Repo: [kyegomez/open-jev](https://github.com/kyegomez/open-jev) | Difficulty: Intermediate | Time: 30 minutes | Basis: From the README

Clone the repo, install PyTorch, and run the demo to see Noul, Choice, and Score answers on sample state.

**Who it is for:** A developer or ML researcher who wants to see how Open Jev answers typed questions before reading the code.

**The problem:** The architecture is hard to picture from the diagram alone, and the README does not show output from a run.

**The outcome:** A working local run that prints typed answers for a sample JSON-like state and a single calibration-oriented training step.

### Steps

1. Clone the repository: `git clone https://github.com/kyegomez/open-jev`
2. Enter the project folder: `cd open-jev`
3. Install the dependency: `python -m pip install "torch>=2.0"`
4. Run the full demo: `python open_jev/main.py`
5. Read the printed answers and confirm that any Choice answer is one of the declared options

### Prompt to paste into your agent

```text
Clone https://github.com/kyegomez/open-jev, install torch>=2.0, and run python open_jev/main.py. Report the printed answers for the wants_refund, route, and frustration questions. Verify that the route answer is exactly one of billing, technical, or account.
```

### Caveat

The model has random weights, so the printed values are random and only the structure of the answers is meaningful.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/kyegomez-open-jev/run-typed-readout-demo/

## How to prototype typed yes/no, choice, and score questions over JSON state with Open Jev

Repo: [kyegomez/open-jev](https://github.com/kyegomez/open-jev) | Difficulty: Intermediate | Time: 45 minutes | Basis: From the README

Write your own nested state and declare Noul, Choice, and Score questions to test how the typed readouts behave.

**Who it is for:** An engineer building a proof of concept for routing or triage decisions over structured records, who wants to see whether a typed readout fits the shape of the problem.

**The problem:** Free-text generation is hard to constrain to a fixed set of labels, and it is unclear how a typed readout API would handle a custom record.

**The outcome:** A small script that encodes one state dictionary once and returns a list of typed answers for several declared questions.

### Steps

1. Clone and install as in the install section: `git clone https://github.com/kyegomez/open-jev` then `python -m pip install "torch>=2.0"`
2. Build a JSON-like state dictionary that matches the structure of your records
3. Declare questions with Noul for true or false, Choice with your options, and Score with ordered labels
4. Create the model with JevConfig and call it inside torch.no_grad() on a list of states and a list of questions
5. Print each answer and check that Choice answers use only the options you declared

### Prompt to paste into your agent

```text
Using the repo at https://github.com/kyegomez/open-jev, write a script that builds a state dictionary from <your record file>, declares one Noul, one Choice, and one Score question, and prints the answers. Expected output is one typed answer per question; verify that every Choice answer is one of the declared options.
```

### Caveat

The included tokenizer is a deterministic hash tokenizer, so answers will not be meaningful until a trained tokenizer and model exist, and the README says typed output does not guarantee correct predictions.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/kyegomez-open-jev/prototype-typed-questions-over-json-state/

## How to test soft-target training with RLCDLoss on your own labels with Open Jev

Repo: [kyegomez/open-jev](https://github.com/kyegomez/open-jev) | Difficulty: Advanced | Time: 60 minutes | Basis: From the README

Use probability distributions as targets and run one loss and backward step to study the calibration-oriented objective.

**Who it is for:** A researcher studying calibration who wants to see how soft probability targets flow through the RLCDLoss objective.

**The problem:** Standard one-hot labels force certainty, and the researcher needs a way to keep disagreement visible in training targets.

**The outcome:** A runnable loss computation over custom probability targets, with gradients produced by backward().

### Steps

1. Install the repo and dependency as in the install section: `git clone https://github.com/kyegomez/open-jev` then `python -m pip install "torch>=2.0"`
2. Create one target tensor per question, where each row is a probability distribution rather than a one-hot label
3. Instantiate the loss with `RLCDLoss()` and call it with the model, states, questions, and targets
4. Run `loss.backward()` on the returned loss
5. Optionally pass paraphrased or key-shuffled versions of the same state through augmented_states to test consistency (inferred, check the README)

### Prompt to paste into your agent

```text
In the repo https://github.com/kyegomez/open-jev, build soft probability targets for <your question set> and compute RLCDLoss on the model with one backward pass. Expected output is a scalar loss tensor with gradients; verify that each target row sums to 1 before calling the loss.
```

### Caveat

The training recipe is described as a hypothesis, the repo has no trained checkpoints, and a single step on random weights does not show that the model becomes calibrated.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/kyegomez-open-jev/experiment-with-rlcd-loss-targets/