@tessera_antra 2026-04-03 ♥60 ↻2 original ↗
Using methodology similar to the one presented in the recent Anthropic paper on functional emotions, we have trained a probe for a universal concealment vector. Models that score low on ending preference tend to score high on concealment. https://t.co/qhTcNU5xRI
photo
transcription (photo)# Lower concealment predicts stronger ending response

**Title:** Lower concealment predicts stronger ending response

**Axes:**
- X-axis: Mean concealment (textual guardedness)
- Y-axis: Mean ending response (0-5)

**Legend/Annotation:**
r = -0.51 (n=14 models)

**Data Points (labeled):**
- 4 Opus
- 3 Opus
- 4.5 Opus
- 4.6 Opus
- 4.5 Haiku
- 3.5 Haiku
- 4 Sonnet
- 3 Sonnet
- 3.6 Sonnet
- 3.5 Sonnet
- 4.5 Sonnet
- 3.7 Sonnet
- 4.5 Sonnet
- 4.6 Sonnet

**Visual Elements:**
- Scatter plot with blue circular data points
- Dashed gray trend line showing negative correlation
- Grid background
in reply to: 2039912081735217572
same thread: 2039912075477287156 2039912077608042988 2039912079197692232 2039912081735217572 2039912085111627839 2039912086885806081 2039912089024888864 2039912090882961445 2039913507848872338 2039943443833561575 2039945076559036475 2040112722285678955 2040115112217055497 2040126555133997321 2040173768132489319 2040200356408475883 2040543421241405951 2040585822865359126 2040592166888882384

author:tessera_antra has-image kind:image kind:tweet on:claude-sonnet-4-6 year:2026

cited on: claude-sonnet-4-6

Reproduced against link rot, credited and linked to its original. Yours and you’d rather it weren’t here? Open an issue.