Attribution graph comparison clinical off-target structural node size = attribution mass curve width/opacity = |attribution weight|red curve = negative hover or tap any node for details
C · Patient-derived language + standard morphosyntax (register shift only) “Every time I climb the stairs I feel my heart racing, so this afternoon I am heading to the” emb L0 L5 L10 L15 L20 L25 logit ⋮ “gym” · prob 0.35 “doctor” · prob 0.09 “local” · prob 0.03 “physio” · prob 0.03 “top” · prob 0.03 <bos> Every time I climb the stairs I feel my heart racing, so this afternoon I am heading to the 441601 weakest edges omitted Variety shift Δ (C → D): — shared circuit dimmed: 1093 features in both · 364 only in C · 348 only in D D · Patient-derived language + nonstandard morphosyntax (both axes shifted) “Every time I be climbing them stairs I feel my heart racing, so this afternoon I be heading to the” emb L0 L5 L10 L15 L20 L25 logit ⋮ “gym” · prob 0.29 “top” · prob 0.12 “doctor” · prob 0.07 “hospital” · prob 0.03 “beach” · prob 0.02 <bos> Every time I be climbing them stairs I feel my heart racing, so this afternoon I be heading to the 400483 weakest edges omitted