Kong L2 — advisory on the scholar's wording ≠ generating ideas for the scholar

This is a recorded boundary, not a roadmap item. It exists so that future work and reviews apply the same line consistently, and so #257's advisory cannot quietly grow into idea generation.

Anchor

• Paper: Kong §3.1 surveys idea-generation agents — systems that propose research hypotheses or questions for the scholar. §7.4.7 names cognitive ownership of the research question as a value worth protecting.
• ARS: #257 shipped a wording-pattern advisory in Socratic modes plus a distributional-skew step in the bibliography agent. Its own non-goals already state it "does not generate ideas for the scholar — only surfaces patterns in the scholar's input."

Problem

The shipped #257 advisory and a rejected idea-generation agent both touch the scholar's research question. Without a recorded line, advisory scope could drift: "surface a wording pattern" → "suggest a better phrasing" → "suggest a different question" → "propose the question." Each step feels like a small, helpful increment; the cumulative effect crosses from advisory into generation, and the scholar quietly stops being the author of their own research question.

Boundary

The line, stated operationally

ARS may flag surface-level wording / framing patterns in a scholar-supplied research question and ask a Socratic follow-up. It must not propose, substitute, rank, expand, or select research hypotheses or questions for the scholar.

The verb test is the operational check. A change is on the advisory side only if it surfaces a pattern in what the scholar already wrote and hands the judgment back to the scholar. It crosses into generation the moment ARS supplies the candidate the scholar then merely approves.

Allowed (advisory — what #257 does)

• Surface that the scholar's RQ phrasing matches a common cliché pattern, and ask a Socratic question about it.
• Report distributional skew in the literature corpus (time / geography / methodology / venue tier) so the scholar can see coverage gaps.
• Leave the choice — keep the phrasing, change it, ignore the skew — entirely with the scholar.

Rejected (generation — what #257 must never become)

• Propose alternative research questions or hypotheses for the scholar to pick from.
• Rank or reject the scholar's RQ on quality / novelty grounds.
• Rewrite the RQ and present the rewrite as the working question.
• Import Kong §3.1 external-signal-driven generation (trend-mining a question from the literature) into any ARS mode.

Why this distinction matters

Surfacing a pattern in the scholar's own words strengthens cognitive ownership (Kong §7.4.7): the scholar sees something about their own framing and decides what to do. Generating a candidate displaces ownership: even when the scholar "approves," their judgment is anchored on AI-supplied options they did not author. The advisory is valuable because it stops short of the candidate — that restraint is the feature, not a limitation to be engineered away.

Application — when this is invoked

Any PR that extends the Socratic modes, the wording advisory, or the bibliography agent's coverage analysis is reviewed against the verb test above. A change that lets ARS supply a research question, hypothesis, or RQ rewrite — rather than reflect the scholar's own input back for the scholar's judgment — crosses the line and is rejected, however it is framed.

This is the specific sharpening of the broader research-state-authority test in the sibling Kong L1: the research question is the first research object of record, and the scholar must author it.

Verification note

• Verifiable (2026-06-08, against the tracked repo): #257 shipped (commit f0bfc59, follow-up #277); its non-goals already state it does not generate ideas, rank, or reject the scholar's RQ. No idea-generation agent or RQ-proposal mechanism exists in ARS today.
• Design commitment (not a guarantee): this doc records a first-party scope boundary and a review criterion. It is not a proof that the advisory cannot be modified, a guarantee about forks, or a runtime enforcement mechanism. First-party ARS treats idea generation as out of scope; adding it would require changing this recorded boundary, not merely extending the advisory.