Agent-drafted interpretation, pending human review

The strongest safe conclusion from Apr1 is about normalization, not token weighting. The paired normalization contrast is positive on router probe, and the broad heatmaps show that probe quality varies much more than answer accuracy across the seven-loss menu.

The collection also reinforces two recurring family patterns. First, posterior-side probe quality is often stronger than router-side probe quality. Second, analogical agreement tends to track router probe quality rather than answer quality, which makes it a useful secondary diagnostic for latent strategy structure.

The twr1 bug sharply limits what should be taken forward from this collection. Apr1 still matters as a diagnostic waypoint, but later collections should treat it as evidence about normalization, probe-versus-accuracy separation, and analogical-agreement behavior rather than as clean evidence about token-weighted reconstruction.