When the motif returns

Compare what the source permits us to read with what two providers report under three evidence policies.

Eight selected cases across four albums. Close readings are authored interpretations, not reference labels. No lyrics or model-generated literary prose are reproduced here.

Choose a return

Open collected source record ↗

Earlier target · relevant context · specified return. The middle card can refer to the returning block itself; these are not three successive events.

Change the evidence, keep the target pair

The columns use different input budgets: two retrieved blocks, two selected endpoint blocks, or all retained song blocks. Blocks exclude one-word lines and section headings; this is not the unabridged lyric. Select a cell to inspect its exposed sources.

Provider	Dense top 2	Selected endpoints	All retained song blocks

Red border: failed exposed-citation or target-binding contract. A raw label is still visible; it is not a validated interpretation.

Text suppliedRequired earlier/return target

Where the blocks resemble one another

Embedding size

Block-to-block cosine in the original 1,024-dimensional space. Orange outlines locate the target pair. Hover a cell for its value. Similarity does not establish agreement, identity of voice, or a narrative event.

Executed exploratory revision, October 16, 2026. There are 42 distinct input conditions behind 48 policy/provider cells; identical contexts may reuse a cached completion. These are not independent repetitions. The graph checks source exposure; no model vote adjudicates literary meaning. Reversing the order of already embedded blocks preserves their mean vector, so a pooled vector alone cannot recover this sequence.