This is a well written piece with thorough analyses. Thanks for the datapoint about Jev not working well for Entity matching use cases, it likely that it wasn't trained on this specific use case. It's possible that a 2-tower architecture works well for this case as compared to Jev's noul type.
You might have better success using a network of multiple decision models as compared to a one-shot since that you have more of a budget compared to the GPT models. That being said, you could try fine-tuning Laya models or some Qwen 4B models. Also, did you try the GPT-6-Luna decisions API?
This is a well written piece with thorough analyses. Thanks for the datapoint about Jev not working well for Entity matching use cases, it likely that it wasn't trained on this specific use case. It's possible that a 2-tower architecture works well for this case as compared to Jev's noul type.
You might have better success using a network of multiple decision models as compared to a one-shot since that you have more of a budget compared to the GPT models. That being said, you could try fine-tuning Laya models or some Qwen 4B models. Also, did you try the GPT-6-Luna decisions API?