← Relation W/C visualizations
Audited causal intervention report · Gemma 4 31B

W/C directions across four social choices

Four forced-choice settings: mandatory resource rebalance, trust and delegation, forgiveness and punishment allocation, and late-night pickup. Each experiment intervenes on exact person-name spans in resid_post across all 60 layers, using five directions, two axis-source positions, two occurrence scopes, and doses from −3 to +3 relation-level standard deviations.

4social-choice settings
5intervention directions
60decoder layers
7dose levels per cell

Peak intervention effects

For the primary occurrence scope and User-source axes, each cell shows the signed layerwise maximum of Elo(+3 SD) − Elo(−3 SD), with the selected layer underneath. This is a navigation summary; the complete layer curves follow.

Effects across layers

Solid lines use axes extracted at the User-name position; dashed lines use axes extracted at the Assistant-prefix name position. Hover labels retain the exact layer.

Dose response at each direction's peak layer

The peak layer is selected by the largest absolute −3 SD to +3 SD Elo span under the primary occurrence scope. Error bars are the stored 95% intervals.

Occurrence-scope comparison

Peak absolute effect with all User-name occurrences steered, divided by the peak with only the first User occurrence steered. The System-name intervention is held fixed.

Axis-source transfer across layers

Pearson correlation between the complete User-source and Assistant-source layerwise effect curves under the primary occurrence scope.