1. Semantic proximity
Cosine similarity compares the vectors of two complete text blocks. Here it helps locate echoes; it does not decide whether their claims are compatible.
One motif, several questions
Close in space, close in meaning?
Two passages can share vocabulary, characters and a situation while making incompatible claims. They can also use different words to maintain the same position. Reading these curves requires returning to the proposition: what is claimed, who claims it, and when it applies.
2. Relations between ideas
To identify opposition, we need to fix the referents, scope, time and modality. A different participant or date does not, by itself, create a contradiction.
3. Agreement between readers
When two providers assign the same label, that shows consistency between their readings. It neither replaces a justification nor makes the label a literary truth.
The songs
Follow the echoes without mistaking them for a position
The reference stays fixed: the first selected passage in each pair. Each point compares one song block with that reference in 1,024 dimensions. Both curves share an axis, but the models retain their own similarity scales.
text-embedding-3-small · 1,024 dimensionstext-embedding-3-large · 1,024 dimensionsSelected pair
Cosine similarity curves relative to the first passage. Each block is a button that can be reached with Tab and selected with Enter or Space. The same values are available through the numbered buttons below.
Hover over a point, tap it, or select a block with the keyboard. The gray background marks blocks outside the selected pair; the lines follow block order only. On mobile, scroll horizontally to see the full chart.
A dip means lower similarity to the anchor; it does not establish opposing ideas, an emotional change or an event. The anchor has a cosine similarity of 1 with itself by construction.
Blocks contain up to six retained lines; single-word lines and headings are omitted. Blocks do not necessarily align with verses, individual speaking turns or complete propositions.
Changing the context
Where do the readers agree, and where do they differ?
The pair stays the same; the supplied evidence changes. Each card compares OpenAI and Anthropic labels for relation, voice and event. Agreement is highlighted separately for each field: a single agreement label would hide meaningful differences.
Same labelDifferent labels
The event field has a design problem
The original schema required a choice among actual, hypothetical, quoted, recalled and none. It mixes two questions: whether something happened and how it is presented. An event that happened can also be quoted or remembered. A difference in this field may therefore reflect a different interpretation, overlapping categories, or both.
With all retained blocks, the readers agree on 8/8 relation labels, 7/8 voice labels and 2/8 event labels. That contrast calls for revisiting the questions as well as the answers.
These are eight cases selected after the pilot, with two different context budgets. The displayed labels pass the source-address and target-pair binding checks; those checks do not verify that the text supports the interpretation. These are neither accuracy measurements nor independent tests of generalization.
A smaller, more explicit control
Separate the topic from what the sentence claims
All the texts can be read here: 56 original sentences in four families, each with Spanish and English versions. Every template compares a base sentence with six variations. The negation is intended to be incompatible with the base when referents, time and scope remain fixed; changes in participant, time and modality require different questions.
Base sentence · reference for all twelve points
text-embedding-3-smalltext-embedding-3-large
Six variations of the base sentence, in a fixed order. Each row compares cosine similarities from two models. Use Tab to reach each transformation and Enter or Space to show its text and intended relation.
On mobile, scroll horizontally to see the points. The complete texts and values are also in the table below.
| Transformation | Original control sentence | small | large |
|---|
What this control supports
In three of the eight templates, small ranks the negation above the paraphrase; large does not do so in these templates. This reveals a specific sensitivity to phrasing. It does not establish an overall winner: the eight templates share four families, and negation changes fewer words than paraphrasing.
Lexical overlap has not been controlled for, and no threshold has been fitted. A high score for a temporal or modal variation is not necessarily an error: it preserves the topic while changing the time or strength of the claim. The control helps formulate a question that distance alone cannot answer.