Research log

technical reports from ongoing experiments

• What base models say before the assistant arrives · Aug 2, 2026
  A 19-character probe finds the helpful-assistant persona already in base models whose training corpus postdates ChatGPT. One literary voice surfaces on all eleven models tested.
• What crosses between two language models · Aug 2, 2026
  A linear map matches sentences across independently trained models at 94%, and lineage does not predict which pairs work. No activation transfer beat plain text.
• Two covert channels between AI agents · Aug 2, 2026
  Text monitoring catches the tool-call channel. Activation probes catch the token channel at AUC 0.945 to 1.000. Neither catches both.
• Language-specific neurons in a 284B model · Aug 2, 2026
  Language-tagged cells are over-represented in the last four of 43 layers by 1.9x to 7.0x, at every threshold tested.