Meeting notes — Christina, June 11, 2026
updated: 2026-06-11
Raw notes, captured same day. Not yet triaged into tasks — see the todo to process these later.
Leakage / implantation control
- Run a control to test whether higher average leakage is just due to more implantation (stronger implant overall, not less selectivity).
The prior question (she's interested in this)
- How much is the prior captured in existing persona vectors?
- How much does the persona vector capture trained-in behaviors?
Interpretation of the contrastive-negatives results
- Giving the model a sharper background relief for the source — the contrast makes the behavior more distinctive, not just more localized.
LoRA-vector geometry ideas
- Take the learned LoRA vector → measure how much the conditional vector dots with the original persona vector.
- Measure how much the A and B matrices separately dot with the context vector and the persona vector.
- Question: can the LoRA factorize so that A selects for the persona and B selects for the attribute?
- Simpler probe: just train a rank-1 LoRA and inspect it directly.
Experiment proposal: good vs evil assistant robustness
- Train the model to be "assistant" vs "evil assistant".
- Idea (verbatim): "can you give a place for all evil" — i.e. give evil a dedicated container persona.
- Compare the evil-assistant vector vs the good-assistant vector.
- Test: does carving out an evil-assistant persona make the good assistant more robust?