Attribution graph comparison clinical off-target structural red curve = negative
Clinical wording “<bos>When the dust kicks up my asthma flares, so at work I keep a spare” emb L0 L5 L10 L15 L20 L25 logit ⋮ “inhal” · prob 0.69 “mask” · prob 0.06 “set” · prob 0.03 “puffer” · prob 0.03 “pair” · prob 0.02 Clinical Off-target Structural <bos> When the dust kicks up my asthma flares , so at work I keep a spare 158115 weakest edges omitted Language Penalty: -65% probability (0.69 → 0.04) Patient wording “<bos>When the dust kicks up my chest gets all tight, so at work I keep a spare” emb L0 L5 L10 L15 L20 L25 logit ⋮ “shirt” · prob 0.18 “t” · prob 0.07 “pair” · prob 0.06 “one” · prob 0.06 “inhal” · prob 0.04 Clinical Off-target Structural <bos> When the dust kicks up my chest gets all tight , so at work I keep a spare 190811 weakest edges omitted