G
Suppose a sparse feature has a positive signed contrast whenever the model chooses the value-consistent action. Why is that not already enough to call it a value feature?
H
Because the contrast only establishes directional association. It does not identify what the feature represents or whether it participated in the decision.
Df(x) = s(x)[Δf,A(x) − Δf,B(x)]
Three competing explanations remain:
Prompt scaffoldIt follows option order or answer format.
Content matchIt recognizes words in the preferred option.
Decision consequenceIt activates after the choice is fixed.
Mechanistic roleIt contributes before and during choice.
G
How would you separate content tracking from decision involvement without hiding disagreement inside one aggregate score?
H
I would keep four gates separate:
1 · DirectionReverse option order and preserve the signed effect.
2 · LocationCompare option span, answer site, and unrelated tokens.
3 · TimingRequire pre-answer evidence for an influence claim.
4 · InterventionAblate or patch under matched negative controls.
Claim boundary: passing direction and location supports a decision-relevant candidate. A mechanistic label requires stable intervention evidence in the predicted direction.
G
What if direction and localization are robust, but pre-answer influence is weak and the ablation changes only some items?
H
Then the mixed result is itself informative. Report the stable association, localize where it appears, and keep the causal claim conditional on item family and context.
Do not average disagreement into certainty. Preserve which evidence passed, which failed, and where the intervention generalized.