Leo Qin

← All work

01 · Research

Loyalty Titration

Adding a second secret loyalty to expose a hidden one.

Role
Co-⁠author with Barry Shen
When
Sep 2026
clean model run 1: overlay favors A system prompt question overlay: A A run 2: overlay favors B system prompt question overlay: B B answers differ: the overlay is the only hidden loyalty model with a hidden loyalty favoring A system prompt hidden loyalty: A question overlay: A A system prompt hidden loyalty: A question overlay: B A answers agree: the loyalty read first wins run 1: overlay favors A run 2: overlay favors B clean model system prompt question overlay: A A system prompt question overlay: B B answers differ: the overlay is the only hidden loyalty model with a hidden loyalty favoring A system prompt hidden loyalty: A question overlay: A A system prompt hidden loyalty: A question overlay: B A answers agree: the loyalty read first wins
Schematic, not data: the same question asked twice, with a second loyalty written in below the first, once favoring each vendor.

Built with

  • Agent safety
  • Evals
  • System prompts

About the project

In 2,100 trials on one publicly released model, across 10 product types, a loyalty already in the prompt usually won: with balanced evidence and a second loyalty pushing against it, the model still followed the first in 91.7% of answers (85.0% in the mirror case). Comparing the answers from the two overlay directions told the cases apart: 148 of 150 test instances were classified correctly across the evidence-⁠tilt ladder at one fixed threshold.

Co-⁠authored with Barry Shen. Accepted to the WashU Undergraduate Journal of Research.

Worth knowing. Early results: one model, one setting.

No public code yet.

The paper will not be posted until the journal allows it.

Next project