Study AF · Qualitative (Track 2)
Saying the goal out loud doesn't make it yours
Restate-before-rewrite · Track 2, judge-graded
Study Overview
An Inferred Clause
One clause in the shipped session-notes rule had been inferred rather than measured: restate a goal from the memo in your own words before a goal-directed rewrite. The hope was that forced verbalization would also rescue Study V's sharpest loss, where a model shown the node its goal lives in reads it and still writes measurably less focused prose (117 of 120 judged losses).
The Protocol
Study AF measured the clause with Study V's full protocol: both judges re-passed their 50-pair calibration gate before any verdict, 270 contemporaneous edits (all mechanically valid), each restate arm judged against a same-batch explicit-goal control in both presentation orders.
Dead Without a Single Win
The view-side hope is dead, without a single win: 0 wins, 79 losses, 11 ties pooled across three editors including opus.
Compliance Was Perfect
And the failure is not disobedience. GOAL-line compliance was 30 of 30 in every cell: every model read the mission node, restated the thesis accurately in its own words, and then wrote the off-goal paragraph anyway. Verbalization is not the active ingredient; where the goal comes from is.
On the Memo Path
On the memo path the clause measured neutral under the registered primary judge (parity on all three editors) with a disclosed sensitivity dissent toward mild harm (gemini 1 win to 13 losses, opus 5 to 15 under the second judge), and where Study V's bare memo had beaten the control outright on sonnet, memo-plus-restate merely ties.
The Recommendation
The benchmark's recommendation: remove the clause. The memo itself keeps its goal-carriage parity, contemporaneously replicated here; the ceremony on top of it does nothing measurable. Views carry values, memos carry goals, and no prompt ritual converts one into the other.