Push brackets, pop them, or generate a conflict. The stack reader answers with the true top; the bag reader sums owed closers with no order and answers with the majority (leaning ] on ties — the fitted intercept). The model row applies the fitted readout, gap ≈ −2.33 + 0.77·top + 3.53·bag pre-finetune, −0.17 + 6.00·top + 1.34·bag after (a schematic operationalization: top = ±1 by identity, bag = the owed-closer imbalance). The conflict lamp marks prefixes where the top disagrees with a strict majority (ties have no majority and are not conflicts). On decisive majorities the pre-finetune model visibly sides with the bag against its own stack; on thin majorities the terms nearly cancel — and that near-cancellation, not decisive wrongness, is why measured conflict accuracy is exact chance (0.47–0.50 at depths 3–6). Post-finetune, the same prefixes read correctly at 0.93.