Kong #257 Idea Diversity / Coverage Gap Advisory

Issue: #257 Parent meta: #255 Related architecture boundary: #134 active conductor

Problem

Two low-autonomy advisory surfaces are missing:

1. Socratic research-question formulation can accept generic wording shells such as "exploring the impact of X on Y" without showing the user that the phrasing itself is common and low-specificity.
2. Lit-review mode can fill topic gaps through uncovered_topics while still returning a source set concentrated in one period, site, method, or venue family.

Both signals are advisory. Neither should generate ideas, rank the user's idea, block output, or displace the scholar's choice.

Socratic Wording Advisory

The two Socratic mentor agent copies now carry ## Wording-Pattern Advisory (Kong #257):

• deep-research/agents/socratic_mentor_agent.md
• academic-paper/agents/socratic_mentor_agent.md

The section defines 20 surface-pattern families (WP01-WP20) and the WORDING_PATTERN_ADVISORY output tag.

Important boundary:

• The check is on wording and framing only.
• The check is not a novelty judgment.
• The check does not rewrite the research question unless the user asks.
• The check asks one Socratic follow-up so the user can choose domain-native phrasing.

Lit-Review Distributional Skew Advisory

The lit-review agents now carry ## Distributional Skew Advisory (Kong #257) or the equivalent Step 4.6 section:

• deep-research/agents/bibliography_agent.md
• academic-paper/agents/literature_strategist_agent.md

The pass runs after retrieval/screening/deduplication and before final report emission. It extends the existing uncovered_topics / search-fills-gap machinery with four dimensions:

• time distribution
• geographic distribution
• methodological distribution
• venue tier distribution

Threshold: emit DISTRIBUTIONAL_SKEW_ADVISORY when a single known value accounts for >= 70% of known entries for that dimension. The denominator is known_N for the dimension, not total source count.

Important boundary:

• Missing metadata is not inferred.
• The advisory does not downgrade sources.
• The advisory does not block bibliography or literature-search output.
• The user can keep the skew when it matches the RQ.

#134 Boundary

134 active conductor is a routing and task-envelope layer: it decides what mode should run and how cross-phase work is dispatched. #257 is content-surface advisory:

• Conductor: "What mode should we be in?"
• Wording advisory: "Does this proposed RQ wording resemble an AI-typical shell?"
• Distributional skew advisory: "Does this retrieved corpus over-concentrate on one known distributional value?"

These can coexist. #257 does not add routing logic, autonomous search expansion, or deterministic PreToolUse enforcement. The relevant #134 design context is documented in docs/design/2026-05-18-ars-v3.9.2-phase-boundary-spec.md#relation-to-134-v310-conductor.

Calibration

The repository includes a small hand-curated gold set for the Socratic wording advisory:

• evals/gold/rq_framing_patterns/gold_set.json
• scripts/check_rq_framing_patterns.py

Gold-set shape:

• 40 RQ framings
• 20 wording-cliche positives
• 20 domain-native negatives

Acceptance thresholds:

• FNR < 0.30
• FPR < 0.20
• balanced accuracy >= 0.75

The checker is intentionally lexical and conservative. It exists to prevent over-warning; the prompt-layer advisory remains scholar-facing and non-blocking.

Example

deep-research/examples/idea_diversity_coverage_gap_advisory.md shows both advisory surfaces:

• WORDING_PATTERN_ADVISORY on an "impact of X on Y" RQ shell
• DISTRIBUTIONAL_SKEW_ADVISORY on a lit-review corpus concentrated in post-2023 NLP venues