This work compares two readout strategies while keeping the model checkpoint and prompt fixed:
- Scoring answer tokens directly (scored readout).
- Generating an answer after a written rationale (generated readout).
The experiments span four retail tasks across three markets, yielding 13 model‑domain cells; two of them rely on fully public data and checkpoints. Scored readouts rank outcomes more accurately in 12 of 13 cells (two‑sided sign test p≈0.003), improving AUC by 1.5 to 14.5 points. Paired bootstrap confidence intervals exclude zero for every newly measured cell.
The performance gap varies with task‑specific supervision and the mismatch between training and serving formats: an untuned base model shows a -2.2 point drop, whereas rationale‑format supervision yields up to +13.7 points. Analysis of roughly 9,000 rationales reveals two correlates—reduced reliance on the dominant predictive feature and convergence toward stock formulations. Probability saturation does not explain the gap.
A third readout elicits a probability before any verdict. It markedly improves calibration (Brier score from 0.47 to 0.15) while ranking within the noise of scoring, but only for outcome rates seen during training; it underperforms scoring when the scored head is already calibrated.
We interpret these differences through the objectives each readout aligns with, identify training choices that narrow the gap, and propose retaining generated rationales while sourcing ranking from the scored head.
Review