Evaluation
Evaluation
40 held-out synthetic incidents with planted traps: decoy changes, silent roots, look-alikes and non-change causes. Weights were tuned on a separate 40-incident set, then frozen. Synthetic data, so read this as a demonstration, not a benchmark.
Right change ranked first
—
Cited paths that exist
—
every edge re-checked by code
Simultaneous pages → pages sent
—
Brain answers correct
—