Limitations

Where ControlSift shouldn't be trusted, even after Gemma results land.

Synthetic data Generator assumptions may not match real audit evidence distributions.
Rule-derived labels Not fully human-labeled gold. Challenge uses structural audit + a 20-case spot-check.
Compositional research surface (v1.1) Dual sections, decoys, and SCOPE scaffolding mean strong scores may reflect schema reading, not naturalistic binder judgment.
Uneven class difficulty Classical CONTRADICTORY remains easier than SUFFICIENT / IRRELEVANT.
Frozen prompt Zero-shot doesn't tutor the packet grammar; discovery failures are valid outcomes.
Single small model Gemma 3 1B results don't automatically transfer to other sizes or families.
Single-artifact cases v1 isn't a multi-document evidence binder task.
No live inference product Static research artifact only.
Parse failures Malformed outputs become UNPARSEABLE and count against performance.