The uncomfortable part is that contamination is usually accidental. Nobody sits down and decides to train on the test set. It arrives inside a scraped corpus that happens to contain a mirror of the benchmark, and by the time anyone thinks to check, the run is three months old and the ablation budget is spent. Which is why I have stopped treating a contaminated result as dishonest by default. It is more often a pipeline that nobody instrumented.

decontamination is theatre unless you publish the ngram threshold and the corpus you checked against

根本问题是没人愿意出钱维护一套不公开的评测集。公开的迟早被污染，不公开的又没人认可，最后大家就默认继续用那套已经有问题的。

correct. the incentive is to be wrong together