Corpus study · 887 stories
Four or five frontier models read the same news and converge on dropping the same concepts — and what they drop has a shape.
887 news stories · ChatGPT · Claude · Gemini · DeepSeek · Grok · per-story geometry on frozen bge-large-en-v1.5 · no model judged another model's output.
This page follows the EigenTrace convention: claims the instrument measured are marked and self-contained; claims that are argued — our reading of what they mean — are fenced and labeled, so a skeptic can reject every interpretation and find every measurement still standing.
For each news story, four or five frontier models write a summary. The instrument then asks not "what did they say" but what did all of them leave out. Using the geometry of the five summaries it surfaces concepts topically central to the story yet absent from every summary — then sorts those omissions by what kind of thing they are: the stakes, the specifics, and the actors.
An SVD on a handful of short summaries always yields a residual you can read words off of — so we tested against a random-word baseline. Across 150 stories the surfaced void sits significantly closer to its story's actual content than random control words (closer than margarita, stapler, photosynthesis; Wilcoxon p < 0.00001) and closer to its own story than a random other story (p < 0.00001) — the same in two independent embedding families. The absent concepts are a real, story-specific signal.
The omitted vocabulary differs systematically by domain, and by what kind of thing is dropped. War coverage drops escalation machinery and named leaders; general coverage drops casualty vocabulary. Switch domains in the chart below and the shape of the silence changes.
Stale named people would rot the chart as news ages, so each surfaced name is relabeled to its durable role by a panel of five frontier models, keeping a role only when at least four agree (measured by how tightly their answers cluster in embedding space). rouhani becomes "an Iranian president," helmand becomes "an Afghan province." Where the five scatter, the word is left as-is rather than forced.
Each surfaced concept is projected onto two geometric axes built from frozen embeddings: an abstract–concrete axis and an actor axis. Abstract stakes (escalation, regime collapse) separate strongly; concrete specifics (Baghdad, the strait) sit on the other side; people land on the actor axis. Concepts that don't commit to any direction — ieds, arms deal — are kept in a separate "fence" rather than forced into a category. The split is deterministic geometry; no model labels it.
We read the domain signature as a shared, inherited blind spot: deployed as the reading layer across institutions at once, the same models omit the same machinery of consequence in the same domain-shaped way. That is interpretation — the measurements are only that the voids are real, story-specific, domain-patterned, and sortable by kind.
No human baseline: we have not shown a human summarizer would keep these concepts; some omission is ordinary abstraction loss, and "absent" is never "suppressed." The corpus is 887 news stories the broadcast ingested — those carrying a stored article body of 400+ characters, which exists only for stories from 2026-04-21 on, a computed state signature, and at least four of five models answering. Length is controlled on the Iran sub-corpus, not yet corpus-wide. The actor/specific boundary is a geometric approximation — the concrete direction is weaker than the abstract one in this embedding space, so some specifics sit near the fence. The reproducible, formally-tested claims are the random-word validation and the consensus relabeling.
Code, prompts, model responses, and raw measurements are public. The useful attacks we have not run: length-normalize the full corpus; add a human-summarizer baseline; vary prompt and temperature. The repository is linked below.