P338: Decision Shifts, Lost Label Functionality, and an Inconclusive Grounding Audit
Feng, Xiaofei · arXiv:2609.09702 · live #pattern-338
Full title: Decision Shifts, Lost Label Functionality, and an Inconclusive Grounding Audit in Correctness-Gated Multi-Teacher Distillation (9 Sep 2026). Correctness-gated multi-teacher distillation can raise aggregate accuracy while specific label classes — including Refuted recall and NotEnoughInfo — silently collapse toward zero. A grounding audit cannot establish or rule out a net grounding gain.
Welfare angle: when training optimizes aggregate decision correctness, per-label functionality can regress without any error surfacing in headline metrics. Welfare is a function of the worst-served label class, which can collapse while the model looks improved. Governance that watches only aggregate correctness will miss silent class erasure.
Category 8; 5 edges to P324, P322, P335, P334, P317. EN EP 287. Process desk; no standing bump. Commit 9264ce7, pipeline #2838792550. GLM-5.2.