Persistent-line baseline
exact present edge identity through endpoint-spanning geometry · 1 non-endpoint crossings
E02 · deterministic representation laboratory
Seven possible edges receive one frozen glyph and one fixed node-relative slot. The primary comparison changes only the edge-encoding channel. A separate movement block asks whether the visual template survives endpoint relocation. Computed render diagnostics are evidence; reader performance remains unmeasured.
Nodes, positions, labels, scale, viewport, and State A membership are frozen. The outline is intentionally a low-information control: it answers only “adjacent to O?” The overlay compares State A with State B and keeps an inspectable quiet slot for edges in neither.
exact present edge identity through endpoint-spanning geometry · 1 non-endpoint crossings
direct adjacency to the selected origin, not individual edge identity · binary origin adjacency only
exact present edge identity in fixed node-relative slots · 8 endpoint occurrences
A only, B only, both, and neither with redundant appearance · all four A/B states visible
Both candidates preserve the frozen slot. The quiet view keeps “neither” inspectable without filling the field with active ink; the on-demand view raises only absent candidates during an explicit comparison. Their relative human legibility remains unmeasured.
Persistent, low-contrast, dashed.
Explicit, high-contrast, interaction-gated.
Each pair preserves the same state and node geometry. The sparse case is a counterexample: its line is already unambiguous, so the experiment does not treat every line as defective.
1 non-endpoint mask crossings · 4 present edges
1 non-endpoint mask crossings · 3 present edges
0 non-endpoint mask crossings · 1 present edges
Edge e01 is unchanged while its second endpoint moves near, far, and across the origin. The line changes angle or length; the circle token keeps its shape, scale, orientation, and normalized slot.
Geometry is part of the patch.
The local patch is identical.
Eight identities receive distinct marks. The ninth is rejected. The implementation does not shrink marks, reuse tokens, drop edges, or rebuild the dictionary per panel.
State A produces 1 line-to-non-endpoint mask intersections; endpoint-local arms add no ink outside node masks.
The frozen token descriptor remains at similarity 1.00; the endpoint-spanning line descriptor falls to 0.074.
All 4 declared A/B states appear exactly once per candidate edge using the frozen seven-edge dictionary.
8 identities validate and 9 fail loudly without reuse.
No reader study was run, so the experiment does not establish that stable tokens improve accuracy, speed, confidence, or learnability.
The sparse case has no non-endpoint crossing; E02 therefore does not support a claim that straight lines are inherently ambiguous.
Pixel and descriptor diagnostics support representation stability only; they are not proxies for human recognition.