The data lead wants to fine-tune next quarter’s model o...

The research lead is correct about the risks. Training on production traces leads to collapse, where the model loses the ability to handle long-tail entities, and selection bias, where the model optimizes for queries it already handles well. To make this work, you must require external verification for the traces, deliberately over-sample rare intents to counter frequency bias, and hold out a fresh human-labeled tail set to measure true performance improvements.