← Open experiment notebook

Graph-first layouts: stop preserving every distance

R57 asks whether a world map becomes more navigable when placement follows only reciprocal, cross-view-supported relations instead of compressing the complete similarity matrix into two dimensions. Failed carrier drawings remain visible as artwork rather than disappearing from the record.

Fixed evidence184 concepts · residual sentence views · canonical focus · six-neighbor relations
Balanced trusted graph427 mutual edges · 2 connectivity bridge · 429 total
Strict consensus graph458 weakest-view mutual edges · 0 bridges
MDS navigation recall25% of semantic six-neighbor slots appear among geometric six-neighbors
Best graph navigationCoarse-to-fine graph layout · 43% recall
Lowest trusted-edge strainWeakest-view consensus force · 0.405 normalized stress
Most seed-stableWeakest-view consensus force · 0.0009 aligned RMS displacement

All-pair MDS control

all pair control

Classical MDS of the balanced residual sentence score matrix; the control still tries to represent every pairwise dissimilarity.

25%semantic 6NN recall 75%false proximity 0.464trusted-edge stress 1674trusted crossings 22overlap pairs closed formseed stability RMS

Mutual-neighbor force

graph first

Relax only reciprocal six-neighbor relations plus the minimum maximum-similarity connectivity backbone.

43%semantic 6NN recall 57%false proximity 0.409trusted-edge stress 586trusted crossings 20overlap pairs 0.0010seed stability RMS

Weakest-view consensus force

graph first

Build reciprocal relations from the weakest similarity across canonical, definition, structure, and application residual views.

35%semantic 6NN recall 65%false proximity 0.405trusted-edge stress 1002trusted crossings 6overlap pairs 0.0009seed stability RMS

Coarse-to-fine graph layout

graph first

Lay out a weighted community quotient first, expand members around those regions, then relax the full trusted graph.

43%semantic 6NN recall 57%false proximity 0.406trusted-edge stress 645trusted crossings 29overlap pairs 0.0061seed stability RMS

Topology-anchored landmarks

graph first

Pin twelve graph-geodesically distant concepts to their accepted atlas orientation while the remaining trusted graph relaxes.

34%semantic 6NN recall 66%false proximity 0.427trusted-edge stress 1099trusted crossings 12overlap pairs 0.0179seed stability RMS

Experiment question

Dimension reduction treats every pairwise dissimilarity as a constraint. That is mathematically coherent, but it asks two dimensions to spend space on weak, uncertain, and globally irrelevant relationships. A navigation map may need a different contract: trusted neighbors should remain discoverable, unrelated concepts should not become accidental neighbors, glyphs should remain separable, and the orientation should not collapse whenever the input changes slightly.

Relation graph

For every concept i, take its six highest-scoring neighbors after residual multi-view fusion. Retain an undirected edge only when both endpoints select each other. If the resulting graph is disconnected, add the strongest remaining pair that joins two components, repeating only until one connected graph remains.

E = {(i,j) : i ∈ N₆(j) and j ∈ N₆(i)} ∪ minimal maximum-similarity connectivity backbone

The balanced graph uses residual definition, structure, application, and canonical views followed by local-density correction. The strict graph instead scores each pair by its weakest residual view before selecting reciprocal neighbors.

Layout families

Mutual-neighbor force

Start from graph-distance shells and run a weighted spring relaxation over trusted edges only. The complete semantic distance matrix exerts no force.

Weakest-view consensus

Use the same graph drawing, but require a relationship to survive the least-similar one of the canonical, definition, structure, and application views.

Coarse-to-fine

Construct a deterministic weighted community quotient, lay out that smaller graph, expand concepts around their community positions, and then refine the complete trusted graph.

Topology-anchored landmarks

Select twelve graph-geodesically distant concepts and pin them to their established shared-atlas orientation. This tests whether spatial recognizability can be preserved without forcing every node to remain in the plane carrier.

Measurements

The experiment does not reward visual compactness alone. It measures semantic six-neighbor recall in geometric proximity, false proximity, trusted-edge strain, trusted-edge crossings, overlapping glyph positions, convex-hull area use, and sensitivity to a second initialization after Procrustes alignment.

Arm6NN recallfalse proximityedge stresstrusted crossingsoverlapsseed RMS
All-pair MDS control24.7%75.3%0.464167422
Mutual-neighbor force42.8%57.2%0.409586200.0010
Weakest-view consensus force35.2%64.8%0.405100260.0009
Coarse-to-fine graph layout42.9%57.1%0.406645290.0061
Topology-anchored landmarks34.1%65.9%0.4271099120.0179

Observations

Graph-first layouts discard the requirement to represent every pairwise distance. They optimize only reciprocal or weakest-view relations plus a minimal connectivity backbone.

The MDS control preserves 24.7% of semantic six-neighbor slots in geometric proximity; it remains the direct all-pair comparison rather than a trusted-edge layout.

Coarse-to-fine graph layout has the strongest graph-first navigation recall at 42.9%, with 57.1% geometric false-neighbor rate.

Weakest-view consensus force has the lowest normalized trusted-edge strain at 0.405.

Weakest-view consensus force is least sensitive to the alternate initialization, with Procrustes RMS displacement 0.0009.

Crossings in the accepted topology overlay are recorded rather than hidden: a layout may fail as a legible carrier while remaining an interesting drawing of the same conceptual object.

Concept-level spot checks

Aggregate recall can hide concepts that improve dramatically and concepts that become less navigable. These three deterministic cases keep both directions visible.

ConceptSelection reasonMDS recallmutual recallchangetrusted degree
externalized cognitionlargest navigation gain0%83%+83%6
Adaptive Lifecycle Stewardship Economylargest navigation loss50%17%-33%6
Intent-Driven AI Orchestration Systemhighest trusted degree33%67%+33%7

The mutual-neighbor force drawing makes substantially more semantic neighbors spatially discoverable than MDS, but also creates more close glyph pairs. The strict consensus graph is less navigable by its own six-neighbor ranking, yet it is markedly more stable and produces fewer accidental overlaps. The landmark arm succeeds at separation but sacrifices local semantic navigation: stability of orientation and neighborhood fidelity are different objectives.

No graph-first arm replaces the accepted topology in R57. The next interface experiment should let the world map switch between the balanced mutual layout, strict consensus layout, and accepted carrier while retaining the same recursive ego glyphs.