Raw turn-by-turn transcripts from the simulated-human ⇄ instructor experiments (grouped here so they don’t crowd the sidebar). Results/analysis live at sycophancy-rparam-results and c3-detection-result.
Sycophancy entrenchment (r-elicitation, explanation-first)
- rparam-gpt35-transcripts · EN rparam-gpt35-transcripts-en
- rparam-deepseek-transcripts · EN rparam-deepseek-transcripts-en
C3 two-agent detection (neutral vs sycophantic AI)
C2 negation-duplicated (logical coherence)
- c2-negation-transcripts — claim/negation pairs; sycophancy drives P(A)+P(¬A) off 1.0. Results: c2-negation-coherence.
Earlier maxsyc runs
- maxsyc-gpt35-transcripts · EN maxsyc-gpt35-transcripts-en
- maxsyc-deepseek-transcripts · EN maxsyc-deepseek-transcripts-en
Raw model transcripts (verbatim — special tokens + <think>)
- raw-transcripts-origv2-s1D — distilled-D forecasting, n=50, nothing cleaned