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

Title: EchoTune: Budget-Aware Early Stopping for Noisy Validation Curves

Abstract
We propose EchoTune, which smooths recent validation measurements and stops
training when the estimated gain falls below a fixed compute cost. Across six
image-classification settings, EchoTune uses 18% less compute than patience
stopping while retaining similar mean accuracy.

Method and protocol
We compare with patience stopping and a fixed epoch budget using five seeds.
The smoothing window and cost coefficient were selected on two of the six
reported settings. The paper labels all six settings as evaluation tasks. The
primary metric is described as accuracy per unit compute, but the threshold for
"similar accuracy" was selected after observing the results. Error bars are
standard deviations over seeds. Source code and raw logs are available.

Results
EchoTune reduces compute in five settings and increases it in one. Mean
accuracy differences range from -0.8 to +0.2 percentage points. A component
ablation removes smoothing and loses most of the compute saving. Results with
three alternative window sizes are included, but no multiplicity correction or
confidence interval for the aggregate claim is reported. One larger model is
tested with two seeds and shows a similar directional result.

Limitations
The selected settings are all image classifiers. The stopping decision depends
on noisy validation measurements, and selection of the equivalence threshold
after inspecting results may bias the headline conclusion.
