saey-11-activation-intervals-spectrum-scoring
IN premise — summaries/2026/08/24/bricken-2023-monosemanticity-chunk-4.md
Created 2026-08-25T02:57:55+00:00
Spectrum-aware human scoring divides the activation range into 11 evenly spaced intervals and scores each interval separately against the overall feature hypothesis.
Summary
Instead of getting one lump-sum pass or fail on a feature, this scoring method slices the full activation range into 11 equal pieces and checks human agreement in each one independently. That granularity lets the system pinpoint exactly where along the activation spectrum the model's behavior diverges from human judgment, rather than just confirming that a mismatch exists somewhere.