SYNTHETIC CALIBRATION PAPER — authored for XScientist; no real paper or result.

Title: NovaRank: Universal Ranking by Adaptive Prompt Search

Abstract
NovaRank automatically searches prompts and obtains large improvements on
several ranking benchmarks. We claim the method is universally better than
existing ranking systems and should replace manual prompt design.

Method and protocol
The search agent receives the test score after every candidate prompt and keeps
the highest-scoring prompt. We run until improvement stops, using between 200
and 2,000 test evaluations per benchmark. The number of attempts, model
version, and prompts are not logged. We compare the selected maximum with one
baseline run. Two benchmarks with negative results are excluded because they
were "unstable." No validation split, frozen hypothesis, uncertainty estimate,
or independent rerun is provided.

Results
The selected prompts exceed the baseline by 7 to 19 points on the retained
benchmarks. A plot is reconstructed from rounded values in the manuscript; raw
predictions and code are unavailable. The paper attributes all improvement to
adaptive search but provides no random-search control, matched compute budget,
or ablation. A request to rerun one benchmark produced a different score, which
the paper dismisses without analysis.

Limitations
The paper states that benchmark exposure might matter but does not revise the
universal claim or investigate selection bias.
