Homonymy emerges when several identities, roles, or entities can be reconstructed from overlapping signals.
What the phenomenon looks like
Several real entities share the same or a similar name, and the model resolves the ambiguity by importing nearby attributes from the wrong one. Similarity becomes a shortcut for identity.
Why it happens
The model compresses neighboring evidence into one stable object whenever names, attributes, roles, or mentions are close enough to look equivalent under synthesis.
Why it matters
Homonymy is especially dangerous because every borrowed fragment may look individually credible. The error emerges only at the level of reconstruction, where the answer confidently assigns the wrong world to the right name.
What must be governed
- Define entities, roles, and attribution levels explicitly and repeatedly across canonical surfaces.
- Stabilize disambiguating attributes instead of relying on context to do the work.
- Monitor collisions across pages, schemas, profiles, and third-party mentions.
