Allow locking of validated schema properties across stages/batches

[jd-coderepos]

Across the schema discovery stages/batches, a mechanism to "lock" properties that are already correct would prevent models from regressing on previously good aspects.

However this feature mandates absolute certainty on the part of the schema modeler that the property and its corresponding structure would need no change at all.

[GaganaB]

Hi,

I had a couple of thoughts on attribute locking that I wanted to share with you.

Tracking attributes across schema changes
I envision it as follows: Each attribute in every schema has a branch parameter that tracks how consistently the attribute appears across schema variants; Either through a threshold or through explicit expert feedback or any other reasonable metric, these properties are locked on to the schema. The branch parameter could be more diverse in the sense of - tracking how many consecutive schemas it appears in and so on.

Compression algorithms to enhance SNR
Another alternate way of trying to lock in properties would be perhaps to use a smaller language model to identify and remove low-information tokens?

Interesting in this case, the mechanism could be appropriate in multiple different stages for multiple different problems.

1. For instance, it could also be useful for providing the initial template for experts to iterate on.
2. This kind of directed attention could help compress the context window while preserving task performance for your issue mentioned here.

Hope this makes sense.