Observed arrival · 2026-10-11
Counterfactual Deception Localization
A research dataset and explorer that estimate where a language-model reasoning trace becomes likely to deceive.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
- For
- Researchers studying deception in language-model reasoning
- Worth noticing
- The homepage describes five environments and a trace format that includes sampled continuations, deception rates, and Wilson intervals.
Field notes
The described workflow holds each sentence prefix fixed and estimates a deception rate from continuations graded against simulator state. Its trace schema preserves the sampled generations alongside boundary positions, valid-sample counts, and Wilson intervals. The page says the data spans five environments and presents filtering by environment, model, juncture position, and outcome; the linked paper is currently marked temporarily removed.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
○OpenPublic substance visible
✦PrettyNotable craft visible
●ProPolished or operationally mature
◎NicheUnusually specific use
One card from the complete issue