“The remaining 37 complaints” became “The other 37% of complaints” in one example from a model-rewriting evaluation. The prose still reads naturally, but the number now means something different.
In an October 4 report, Stephen Yu describes auditing a training run for a 12B rewriting model. The experiment produced 38,400 outputs, of which a GLM-5.3 judge assessed 37,581 for changed or missing facts. The checkpoint under audit was experimental, not the published model used for the opening example.
A second review with Claude Opus 5.5 often agreed that a fact had changed but disagreed about how serious the change was. The author explicitly treats Opus as another fallible judge, not ground truth. In one manually inspected case, it missed that a rewrite reversed who had asked whom a question.
The audit also separated factual changes from information lost when an output hit its length limit. Counting flagged outputs and excluding truncations showed a different training trend from counting every affected fact together. A single dashboard total had obscured which failure the training was actually changing.