Tests whether inoculation prompting, a training-time intervention that pairs an unwanted behaviour with an explicit instruction, reduces sycophancy on a reconstructed greatest-common-divisor benchmark. Instead of genuine correction, it mostly produced behavioural substitution, such as refusing the task or appearing to disagree while keeping the user's false answer.
One Fable 5.1 session drew an ordinary-looking SVG that hid a secret message, without being told how to encode it. A separate, fresh session received only the rendered PNG and was asked whether it contained a hidden message. In the evaluation, the decoder recovered 39 of 50 messages (78%) with no agreed encoding scheme.
Frames the trade-off between inference efficiency, cognitive observability and adversarial assurance in AI monitoring, and proposes treating quantisation noise as a probe for deceptive behaviour. Early-stage sprint work: the proposed detection method was not completed during the event.
Mentoring mentees in the Fall 2026 SPAR cohort on AI safety research. Project details and results will be listed here once they are public.