Supplementary material

Each set contains the target and the first N − 1 context concepts in alphabetical order, so composition and count change together. Paper assignments are recomputed for each set. A concept can therefore receive papers at one N but none at another. Radiation/quantum enters the set at N = 8; spectroscopy and thermodynamics enter at N = 10.

A measurable response requires assigned papers and enough time windows in which both the original and the concept-removed geometry can be calculated. The implemented geometry needs at least two active concepts in a window, and the standardized before-and-after comparison needs at least two valid windows on each side of the pivot. No assigned papers and too few valid comparison windows are different reasons for an undefined response. Neither is treated as a measured zero or included in the ranking.

Special relativity, N = 5: gravitation has no assigned papers. Special relativity and electron theory have assigned papers, but removing either leaves too few active concepts to calculate the post-pivot geometry. The matched comparison retains seven pre-pivot and no post-pivot windows for special relativity, and three pre-pivot and no post-pivot windows for electron theory. Only aether optics and electrodynamics have defined responses; the target itself cannot be ranked.

Special relativity, N = 8: gravitation, instrumentation/measurement, and mechanics/time measurement have no assigned papers. Aether optics and electrodynamics each have one assigned paper, but their concept-removal comparisons retain only one valid pre-pivot window each (and 17 post-pivot windows), which is insufficient. The three ranked concepts are special relativity, electron theory, and radiation/quantum. The label 1/3 therefore means first among these three, not first among all eight configured concepts.

Special relativity, N = 10: gravitation, instrumentation/measurement, mechanics/time measurement, and radiation/quantum have no assigned papers under this configuration. The six ranked concepts are special relativity, aether optics, electrodynamics, electron theory, spectroscopy, and thermodynamics/kinetic theory. The label 1/6 means first among these six. Aether optics and electrodynamics now have enough valid comparison windows, although each still has only one assigned paper: measurability depends on the surrounding geometry and temporal coverage, not paper count alone.

Among ten sampled margin scales, special relativity never ranks first at N = 5 and ranks first at N = 8 only at f = 0.25. These are discrete samples, not a test of every margin scale or every possible concept subset. Gödel and deep learning rank first at N = 8 and 10; attention remains at ranks 3–4.

Jaccard overlap is the number of papers shared by two lists divided by the number of distinct papers appearing in either list. For example, lists containing papers A, B, C and B, C, D share two of four distinct papers, giving an overlap of 0.5. This measures agreement between selections, not the fraction of scientifically correct assignments.

A paraphrase is one of the five descriptions of a concept. We combine descriptions by averaging their embedding vectors, then use the resulting vector to select papers. Orange bars compare one description with another. Blue bars compare one description with the other four combined. Black bars compare two descriptions combined with the remaining three combined; the two groups share no descriptions.

The hollow green bars compare four descriptions combined with all five. This asks how much the selection changes when one description is omitted. Because the two sides share four descriptions, high overlap here is not independent evidence that different descriptions produce the same assignments.

For each concept, we average the overlap over the usable comparisons. Bars show the median of those concept-level averages; whiskers show the interquartile range across concepts. Comparisons in which both lists are empty are excluded because they provide no selected papers to compare. If only one list is empty, the overlap is zero. A concept contributes to a bar only if at least one comparison remains.

Assignment rule used in the paper: a nearest-anchor-group 60th-percentile similarity threshold and a confusion-aware margin $\tau(c)$, with $f_{\mathrm{knee}}$ fixed per case. Only the anchor under study varies; the other nine retain their full five-paraphrase anchors. Thresholds and margins are recomputed for each variant.

The numbers in parentheses in the legend count comparisons per concept: up to 10 single-paraphrase pairs, 5 unordered 1-versus-4 splits, and 10 unordered 2-versus-3 splits. The fractions above bars instead count contributing concepts. For special relativity, these are 1/10 for orange, 6/10 for blue, 5/10 for black, and 6/10 for hollow green. The sole contributing concept for the orange bar is electron theory, not special relativity itself. Differences between these bars therefore combine differences in description choice and concept coverage; they are not a matched comparison across the same concepts.

The corrected comparison excludes empty-versus-empty pairs and counts each unordered split once, regardless of its original orientation. The earlier assignment-rule comparison remains in the downloadable source data; only the assignment rule used in the paper is shown here.