Synthetic fixture / experiment worksheet

Test the vector.

Build one contrastive direction from paired two-dimensional activations, then compare its projected shift on held-out axes with a same-norm sham. The numbers teach an ablation pattern.

Evidence boundary: This fixture does not run an LLM, capture model activations, score generated text, or establish behavioral control. Replace it with real held-out outputs before making a model claim.

01 / Construct

Four matched positive/negative pairs produce a mean difference. Held-out axes are never used to build the vector.

Pair	Positive h	Negative h

Direction v = normalized mean(positive − negative).

02 / Intervene

Gain 1.00

A zero gain must yield zero projected shift. Negative gain reverses the direction. The sham is a perpendicular, same-norm direction.

Direction

03 / Inspect

Mean absolute projection on two synthetic target axes and two off-target axes.

Target projection

Off-target projection

Held-out ablation receipt

The teaching fixture has two target axes and two off-target axes. A real receipt also needs prompt groups, generated outputs, scoring rules, confidence intervals, model fingerprint, hook name, token scope, and layer.

Arm	Target shift	Off-target shift	Claim