That's fine, good even. Afaik at least for some of these tasks dev teams are doing a lot of manual tuning of the model (rumored that "r in strawberry" had been

That's fine, good even. Afaik at least for some of these tasks dev teams are doing a lot of manual tuning of the model (rumored that "r in strawberry" had been "fixed" this way, as a general case of course). The more there are random standalone hacks in the model, the more likely it will start failing unpredictably somewhere else.