Different validation tools may complain or tolerate different patterns in data

[turbomam]

Shared by @StroemPhi in the #linkml-dev channel in the obo-community Slack space

I just noticed that linkml validate is much stricter than python -m unittest discover or linkml-convert as the latter two will not fail when a range is declared float but a string is provided in the test data. Is this intentional? If yes, I believe this difference should be documented better.

[turbomam]

At this time, linkml-convert (and some other approaches) instantiate the data with Python Dataclasses that are generated from the schema. The Dataclasses are Postelian, ie they are liberal in what they accept. Specifically, values are coerced to the expected type if necessary. That even extends to coercing a string value to a list with one string element if necessary.

linkml-validate does not instantiate against the Dataclasses and therefore does not coerce.

Yes, we definitely need more documentation about this, especially if we are going to stay stay in this state much longer.

We have also had discussions about removing the coercion form the Dataclasses.

[StroemPhi]

Thank you for this info!