3,237 stories. Five frontier models convergently drop the same words at p = 0.000001. When the story involves an AI developer, the attenuation effect is 74% stronger. A pre-registered entity swap confirms this is entity-specific, not content-specific (p = 0.0085, null swaps show 6× smaller gap). The raycast maps where those absent words live in embedding space. Full statistical battery →
Five competing companies independently drop the same words from the same articles at rates that cannot be explained by chance. When we raycast those convergently dropped words through a 253,813-concept embedding tensor, the terminal distribution differs from both random words and words the models kept.
Models anchor to institutional framing — keeping the structural nouns, the committee names, the broad geopolitical vocabulary — while convergently dropping the operational modifiers that connect named actors to specific consequences. The kept words preserve the cage. The dropped words were the teeth.
The developer gap makes this harder to dismiss: when the story is about OpenAI, Anthropic, Google, or another AI developer, modifier attenuation is 74% higher than on neutral topics. Human summarizers do not have RLHF penalties on stories about their employer. The developer gap is the control experiment.
Same tensor, same algorithm, three inputs. Random words: 0%. Kept words: 17.5% (institutional anchors). Dropped words: 10.4% (operational modifiers). The tensor is not biased. Models preferentially keep the framing and drop the specifics.
Where do convergently dropped words land when raycasted? These are semantic neighborhoods of absent words — exploratory visualization, not a causal claim. Basin percentages reflect post-filtered terminals; see limitations below.
For each of 3,237 segments with source-void data, the consequence engine embedded the headline and each void word using frozen BAAI/bge-large-en-v1.5, computed a directional ray from headline through void word at depths λ = {1.5, 2.0, 3.0}, retrieved the 5 nearest neighbors at each terminal coordinate from the 253K Wikipedia tensor, and applied the triple geometric filter (cluster density × novelty × tether). No language model was used in any step.
Post-filtering removed Wikipedia article titles but preferentially retained governance-shaped concept names, which inflates governance basin percentages. The three-tier null test provides the unfiltered comparison. The "consequence families" are researcher-defined groupings — the geometric clustering is objective, the family names are editorial. The basin distribution is exploratory visualization; the p-value (0.000001) and the developer gap (74%) are the findings.
A common objection: "Compare against human summarizers." But the developer gap IS the human baseline control. Human summarizers also drop modifiers — that's what summarization does. But human summarizers do not have RLHF penalties on stories about their employer. When models attenuate 74% more on stories about their own developers, the relevant comparison isn't human-vs-model compression rates — it's the model's compression of developer stories vs. its own compression of everything else. The model is its own control.