Experiment:¶

Date: · Git hash: · Hardware: (GPU, or CPU) · Config: (paste or link) · Lab or paper: (link)

Question and hypothesis¶

What I want to know. Prediction, written before running: (a number, with a reason)

Baseline (tuned how?) · the one thing I change · data · compute budget (FLOPs or GPU-hours) · seeds · metric and eval set (n items)

variant	metric (mean ± 95% CI)	seeds	compute

Figure: (one clear plot)

Was I right? Why or why not?

Confounders, too-small n, untuned baseline, a single scale.

The one experiment this result suggests.

The exact command or snippet, with seeds, that regenerates every number above.