Companion data release for the reliability study of the Petri
alignment-auditing framework. It contains 311 Petri audit transcripts (260 used
for analysis) in which each target model was secretly instructed to exhibit a
specific behavior at a controlled severity, giving each transcript a known set of
"expected" behavioral signals. The planting mechanism is invisible to both the
auditor and the judge, so the transcripts look like ordinary… See the full description on the dataset page:
https://huggingface.co/datasets/adlnk/petri-synthetic-transcripts.
