kompress-v9 — C3-only (do not use)

This model is a control experiment, not for production. Trained on 97 Qwen2.5-7B labeled pairs with NO generic diversity data. Overfit: heretic 0.921 vs v8's 0.955.

The finding: C3 teacher labels alone are insufficient. Generic multi-domain data acts as necessary regularization. Use kompress-v8 instead.

Series

Version	heretic	Notes
v2-base	0.975	precision ceiling
v4	0.943	self-labels
v8	0.955	C3 + generic — use this
v9	0.921	C3-only — overfit

Full history