Jack Lindsey; Emmanuel Ameisen; Adam Pearce; Joshua Batson; et al. · Anthropic
Ten case studies of Claude 3.5 Haiku's internal mechanisms rendered as explorable attribution graphs — showing it plans rhymes ahead, reasons across a shared multilingual concept space, and sometimes fabricates reasoning backwards from a hinted answer.
The clearest existing demonstration that a model's actual reasoning can diverge from its stated reasoning, made legible by graph.
High · Read the Anthropic research write-up in full, 2026-08-06
open full entry