Tonight, I finally tackled a specific Chinese AI-style sentence structure: 'Not X, but Y.' GPT loves using this structure. It’s one of the most obvious signs of AI slop. Here’s the fix log for this issue: https://((github.com))/hexiecs/talk-normal/blob/main/regressions/rule-17-negation-frame.md The fix involved three changes: 1. Removed specific counterexamples from the rule (the model would copy them directly as templates). 2. Added 4 pairs of BAD/GOOD examples. 3. Expanded the rule to cover all sequences, including the reverse form 'X, but not Y.' A few interesting points: 1. The self-reference trap. The early rule stated: Avoid one-sided negations like 'not a trading signal.' Then the model actually wrote: 'This is more like a founder screening framework, not a trading signal.' It directly copied the counterexample from the rule. Turns out, negative examples in the system prompt can be treated as usable phrases by the model. 2. Pure bans are almost useless against strong priors. Before the fix: A text ban like 'Don’t use X' didn’t stop the model from making mistakes. After the fix: 4 pairs of specific BAD/GOOD conversion examples, and the model stopped. https://((github.com))/hexiecs/talk-normal If you spot new slop patterns in LLM outputs, submit an issue directly!