The Control You Cannot Run
Entanglement, Confabulation Floors, and What Self-Report Probes Actually Measure
R2 · elicitationApart Research, Digital Minds sprint, August 2026.
- Question
- A 7B instruct model reported noticing something unusual on 66.5% of responses when nothing had been done to it. Varying only the harness moved that rate by 36.1 points. That is the confabulation floor any introspective self-report has to clear before it means anything.
- Method
- Pre-registered detection tests under activation injection: free report, forced choice and detection, each measured against a control.
- Result
- Free report recovered information under activation injection, 15.3% against a 1.4% control, but failed localisation. Forced choice and detection recovered none.
- Implication
- The detection tests were pre-registered and their failure published. A self-report probe means nothing until it is measured against its own confabulation floor and its sensitivity to the harness.
66.5%reports of something unusual when nothing was done
36.1 pointsmoved by changing only the harness