Skip to content
03 · Cover the Motion, Not the Look/Evidence · 23
Home ↗
Slide 23 bridge

From targeting a benchmark to filling training-set gaps

Chapter 3 covers a target’s motion. Chapter 4 diagnoses the training set without a benchmark.

Three panels: chase the target, aim by search, cover the manifold
The reference changes from target motion to gaps in the training set. Bridge experiment: one seed, 2,000 generated pairs per arm.
Which is it? Copying a benchmark's labels per sample gains +2.7 on that target for −2.8 on everything else. The Chapter 3 search made the same move more mildly: +4.5 on target, −3.0 generality, measured. Benchmark-blind hole-filling lands within 1–3 points of every aimed ceiling with nothing sacrificed. Even the deliberate anti-target arm could not hurt the benchmark it aimed away from.