The one distinguishing feature of an ORM is that it drops SQL and the relational paradigm and places the developer completely within the OOP world. If you drop this, you have merely a database connector library, with much more freedom of behavior than an ORM.
Yes, since they are an notoriously bad abstraction, every ORM will give you an escape hatch for the minority of tasks that it can't abstract at all. That escape hatch is not in any way a defining feature of the system.
Now, about that extra freedom that you get from dropping the requirement that your connector is an ORM... well, neither your data at rest, the presentation to the user, the abstraction for a service client, nor your validation rules benefit from OOP. Proof of that is that OOP interfaces from all of those things precede the current popular ones, and all of them were abandoned due to fundamental dissonance between the model and the data.
The rationale for an ORM is that, even though none of the interfaces you actually want for your data is best done in OOP, somehow OOP is still the best way to integrate them so that you can reuse data and metadata. This thought is not completely without merit, but there is very little empirical motivation for it, and the tiny amount that exists is more than completely explained by the OOP hype that only started to die around a decade ago.
EDIT: Oh I saw you wrote SQL Alchemy! First, thank you for that great piece of software.
Now, SQL Alchemy does provide a lot of useful ways to postpone the object representation or even to map your values into dumb data. My comment up there is on theoretical limitations, but on practice, I do think it's the best option available on Python. (And maybe about the best option afforded, since the language is intrinsically biased into OOP.)
Yes, since they are an notoriously bad abstraction, every ORM will give you an escape hatch for the minority of tasks that it can't abstract at all. That escape hatch is not in any way a defining feature of the system.
Now, about that extra freedom that you get from dropping the requirement that your connector is an ORM... well, neither your data at rest, the presentation to the user, the abstraction for a service client, nor your validation rules benefit from OOP. Proof of that is that OOP interfaces from all of those things precede the current popular ones, and all of them were abandoned due to fundamental dissonance between the model and the data.
The rationale for an ORM is that, even though none of the interfaces you actually want for your data is best done in OOP, somehow OOP is still the best way to integrate them so that you can reuse data and metadata. This thought is not completely without merit, but there is very little empirical motivation for it, and the tiny amount that exists is more than completely explained by the OOP hype that only started to die around a decade ago.
EDIT: Oh I saw you wrote SQL Alchemy! First, thank you for that great piece of software.
Now, SQL Alchemy does provide a lot of useful ways to postpone the object representation or even to map your values into dumb data. My comment up there is on theoretical limitations, but on practice, I do think it's the best option available on Python. (And maybe about the best option afforded, since the language is intrinsically biased into OOP.)