The public visualization is not the starting point; it is the current residue of many small mathematical experiments. This notebook keeps the questions, controls, measurements, failed carriers, and unexpectedly beautiful crossings visible so the path from intuition to interface can itself be explored.
Can the embedding geometry become more discriminating without abandoning its useful local neighborhoods?
The balanced arm reached 1.25× raw score IQR, hubness 0.182, and 83% mutual six-neighbor links.
Does full-article evidence remove shared format, or does it need separate residual normalization before it strengthens the signal?
The strongest residual sentence arm reached 2.67× score IQR and 0.782 2D fidelity.
Can reciprocal relation graphs produce a more navigable world than all-pair dimension reduction?
Coarse-to-fine graph layout raised geometric six-neighbor recall to 43%, versus 25% for the MDS control.
Can the graph layouts become directly comparable inside the recursive ego interface while every local identity remains invariant?
The ego world now switches among 6 coordinate fields. The same active array drives the outer roots, every nested stencil, and the corner overview.
Can the recursive ego field expose isolation, direct origin affinity, and contextual path support without abandoning bounded navigation?
The 0.60 cosine lens leaves 43 concepts isolated, with median degree 3 and maximum degree 46.
What do rank, absolute thresholds, reciprocity, direct affinity, path cohesion, weakest-link paths, and multipath support actually look like under comparable presentations?
The same absolute arm draws 0 visible non-root nodes for C040 and 113 for C012.
Can relaxed shape become its own retrieval vocabulary, independent of the claim that any one representation is best?
184 morphologies span 0–409 semantic cables. C124 and C175 reach 0.934 shape similarity at only 0.161 cosine similarity.
Can recurring-edge and A/B set comparison avoid endpoint-spanning geometry without silently changing token identity?
The frozen token descriptor remains at similarity 1.00; the endpoint-spanning line descriptor falls to 0.074.
Every experiment fixes its input snapshot, names the relationships it is allowed to change, keeps provider-derived vectors reusable, and separates quantitative evidence from promotion into the interface. A drawing may fail as data visualization and still remain valuable as an artifact of the concept system.