Rejected candidates (with reasons)

These come from the ChatGPT second opinion and were screened out because they conflict with the two load-bearing design rules of this menu: verdict/map separation enforced structurally, and no unattributed confidence caches.

• An inquiry "control room" that carries the verdict. The proposed inquiry frontmatter holds current_answer and confidence, with body sections "Current Bottom Line" / "What Would Change My Mind." That fuses the verdict into the map — exactly what adjudication as a separate layer forbids, for the mechanistic reason that contradictions loaded into one context get silently averaged. An inquiry/case-root artifact is fine, but it should hold the question, scope, hypotheses, and pointers to the map; the bottom line lives in a linked assessment artifact.
• Bare confidence fields (confidence: low-to-medium, extraction_confidence: high, status: disputed). A credence is not recomputable from ground truth, so as a frontmatter field it is precisely the "stale trusted cache is a trap" failure the mark discipline forbids — see confidence must be attributed, never a mark.
• Implementing the full list directly in the framework. The build order reads as a Commonplace roadmap; the boundary here is build-local-first in epistack-casebooks with collection-local types, upstream only what survives a worked case. Read its build order as a sequencing of experiments in the sibling repo.
• Enumerating ~15 link labels up front. Defer to collection-owned vocabulary (ADR 019) grown from need, guarded by the articulation test — except correlated-with / independent-of, kept in dialectical link vocabulary because they carry the independence-clustering idea.
• Stance-carrying fields on the claim artifact (polarity: supports, status: disputed). These smuggle a verdict into what should be a neutral node; the dialectical register's contract is what keeps the claim proposition decoupled from any verdict. Support/rebuttal belong on edges, attributed to a party, not as a scalar on the claim.