@gerardsans
Original paper: https://t.co/oka6G5cnMB Refutations already in the literature: Gong et al. (2026) showed Patchscopes are unreliable: injected states overridden by model priors (faithfulness drops sharply). The “layer 6 belief” is just a partial vector sum the rest of the pass overwrites. https://t.co/Z2Dxtvgmnt The architecture is unchanged. The interpretive frame drifted. Time to review the math. Apply null hypotheses thoroughly. Leave anthropomorphic narratives behind for good.