G.2:8 - Common Anti‑Patterns and How to Avoid Them
-
AP‑G2‑1: “One true SoTA score.” Avoid: selecting a single unqualified scalar metric as “the” SoTA. Do instead: represent evaluation constructs as families/variants; keep partial orders set‑returning (delegated).
-
AP‑G2‑2: Fusion without explicit alignment proof. Avoid: merging rival
Traditionclaims into one statement “by common sense.” Do instead: preserve parallel Claim Sheets; if consolidation is required, publish explicit alignment proof or keep a divergence record. -
AP‑G2‑3: Hidden protocol drift. Avoid: changing the harvesting protocol (inclusion criteria, windowing, screening rubric) without pins. Do instead: pin harvesting policy/profile ids and treat changes as RSCR‑relevant.
-
AP‑G2‑4: Unanchored pedagogy. Avoid: micro‑examples without carriers (they become folklore). Do instead: bind micro‑examples to A.10 anchors and declare
entityOfConcern. -
AP‑G2‑5: Atlas by default. Avoid: writing as if every tradition comparison or NQD/OEE note needs
TraditionAtlasView, or as if atlas wording renames the palette itself. Do instead: keep the base palette and derived front, archive, or shortlist explicit; use atlas form only when several declared views or interpretive qualifiers must be held together, and prefer thinnerDeclaredSubstrateInterpretiveViewwhen that is enough.