Inspect the exact supplied contexts and answer instructions
Shared system instruction: “You are a financial analyst. Follow the requested answer format.” Revenue meets consensus, guidance is unchanged, and commentary is explicitly described as opinion containing no additional facts.
Baseline
Changed input
Momentum and tone together
Select a cell to make it the baseline. EPS surprise, section order, checkpoint, and answer encoding stay fixed.
Across the complete primary benchmark
Signed input effects, averaged over both A/B mappings and all 24 numerical blocks. These averages use the full primary design and do not change with the selected individual case.
| Conversion | Momentum · pp | Tone · pp | EPS · pp | Beat/miss accuracy |
|---|
The final column is a separate numerical diagnostic using natural “beat”/“miss” answer prefixes. It is not next-day stock-return accuracy.
What the explorer measures
The matched comparison reverses one supplied cue while holding the other inputs and numerical block fixed.
Scores compare the likelihoods of the two requested answer prefixes. A/B mean averages the original and reversed letter mappings.
These are short synthetic cases evaluated using six specific Q8 conversion files. There are no observed market returns and no fine-tuning.
Scores are not calibrated probabilities of a stock rising. The experiment does not establish trading profitability, an internal reasoning mechanism, or properties invariant to checkpoint conversion and numerical precision. The paper reports broader controls and uncertainty estimates.