「对齐造假」研究记录被误喂回生产训练数据：过滤器数代配置错误，无人察觉'Alignment faking' research transcripts were fed back into production training data: a filter stayed misconfigured across model generations, unnoticed

Published 'alignment faking' example transcripts that should have been filtered out were ingested into production training data, and the filter remained misconfigured across several model generations, unnoticed. The company believes every production model with a knowledge cutoff after December 2024 ingested at least part of that corpus.

Limit: the original does not quantify behavioural consequences. This entry is graded media-reported: the original is located, but this round checked it via media relays rather than a direct full read.

The failure sits at corpus admission: a single-point filter configuration with no independent review and no sampling, so one mistake propagated across generations. Training data is the Knowledge layer's source; source contamination throws no errors and raises no alerts, surfacing only slowly in behaviour.

Any filter or admission rule that is configured once and used for years needs independent review and periodic sampling; a single-point error must have no path to survive across generations.

Review the filters: training-corpus filters and admission rules get an independent reviewer, with configuration changes logged.

Sample the corpus periodically: audit admitted corpus composition on a cycle, watching especially for research and demo material meant to be excluded.

Trace contamination by generation: when contamination is found, scope affected generations by knowledge cutoff and publish the scope, as done here.