Hypothesis: Placing the objective at the end of a prompt produces more accurate responses than placing it at the beginning.
Method: 200 prompts, two orderings each, scored by a separate evaluator model on a 1–5 rubric across three dimensions: accuracy, completeness, coherence.
Results:
| Placement | Accuracy | Completeness | Coherence |
|---|---|---|---|
| Objective first | 3.8 | 4.1 | 3.9 |
| Objective last | 4.2 | 3.9 | 4.1 |
Interpretation: Objective-last shows a modest accuracy gain (+0.4) at a small completeness cost (−0.2). Net positive. The effect is consistent across domains but strongest in multi-step reasoning tasks.
Next: Replicate with a different evaluator to rule out evaluator bias. Run n=500.
This is Experiment 001. I will publish one per week until the pattern breaks or I run out of hypotheses.