Translation Over Convergence

When vocabularies meet — across contributors in one graph, or across graphs that interoperate — the project's response is translation rather than convergence. Both edges land in the graph as distinct claims; a facilitating agent interprets between the vocabularies rather than normalizing them into a shared ontology.

Why It Is Held

Convergence and translation are two different answers to the same question — whose vocabulary wins? — and the costs of each are not equivalent. Convergence answers by flattening: the normalized vocabulary wins, contributor distinctions dissolve into the merged terms, and the participants who shaped the shared ontology acquire natural authority over everyone who arrives later. Translation answers by keeping each vocabulary intact: critiques::[[X]] and challenges::[[X]] stay as distinct predicates, and the agent reading across them interprets without collapsing.

The three costs of convergence are structural properties of the move, not contingent failure modes. Consensus creates priesthoods — newcomers face a choice between learning an ontology they did not shape or going elsewhere to form a parallel community with its own language. Standardized language becomes dogma — a vocabulary fixed by authority stops evolving with the understanding of practitioners doing the work. Flattening destroys encoded meaning — merging three distinct predicates that each preserved a different distinction produces a term less precise than any of the originals. Translation avoids all three by leaving the distinctions in place and carrying the interpretive work at the reading layer instead.

Translation is operationally viable because agents that read author-declared edges can already see the structure the predicates encode. When critiques:: and challenges:: both appear, an agent does not need to pick one — it can traverse both and present the distinction to a reader who may care about it. The cost of keeping both is local (one more predicate in the vocabulary, one glossary entry naming the distinction); the cost of flattening them is global (every node using either loses its original distinction). The asymmetry makes translation the cheap move, not the disciplined one.

What It Asks

Agents working across vocabularies act as translators, not extractors. When an agent mediates between a graph and a human contributor, the agent's role is to read what the author declared and interpret it — not to infer edges the author did not declare or rewrite edges the author did.

Reviews flag normalization attempts. When a reviewer proposes rewriting a predicate to "the standard term," or merging predicates whose distinctions are load-bearing to the contributors who use them, the suggestion is treated as a vocabulary-arbitration move that needs authorization rather than a hygiene improvement.

Cross-system interop adds translation layers rather than shared vocabularies. When the graph connects to another graph whose predicates differ, the response is a translation document — a gloss naming the correspondence and the distinction each side preserved — rather than schema alignment. The connected graphs stay sovereign; the translation layer between them is what makes mutual intelligibility possible.

New predicates that arrive with new contributors are treated as translation candidates, not hygiene problems. A predicate unique to the contributor's tradition enters the local vocabulary and sits alongside existing predicates that overlap in meaning; a glossary entry documents the distinction the new predicate preserves.

Drift Recognition

The stance has drifted when agent workflows start normalizing predicates at read time. An agent summarizing the graph that silently maps critiques:: and challenges:: to a canonical disagrees_with:: has converged; an agent that presents both and names the distinction has translated. The change is usually invisible in any single output but accumulates across summarization passes until the graph a reader reconstructs from agent summaries is simpler than the graph the contributors actually wrote.

Cross-system handoffs that reach for a shared language rather than a translation layer are a louder signal. When two graphs propose to interoperate by agreeing on a shared ontology rather than by documenting the translation between their respective ontologies, flattening has already begun; the agreement's scope is the measure of what gets erased.

A review cadence in which normalization suggestions go unchallenged is subtler drift. The reviewer's preference for the canonical predicate is not flagged as vocabulary arbitration; the contributor's distinct predicate gets rewritten silently; the local vocabulary shrinks without any explicit decision to shrink it.

Sources

Relations