What this is
Word Duel is a semantic implementation of the DGIGM model. Two players take turns choosing words from a shared, finite vocabulary. Every word used is permanently consumed — it cannot be played again by either player.
The game is not about vocabulary size. It is about what happens when the space of available meaning shrinks irreversibly with every turn.
The same structural dynamics that govern geometric constraint in Ω-TRACE govern semantic constraint in Word Duel.
The rules
Two players receive a shared vocabulary of words. Each turn, a player must choose a word that is semantically related to the opponent's last word. The chosen word is consumed: removed from the vocabulary permanently.
A player loses when they cannot find a semantically valid word in the remaining vocabulary. The constraint is progressive, irreversible, and global — exactly as in any DGIGM.
The three axioms are satisfied by construction:
Finite Alphabet (bounded vocabulary),
Irreversible Consumption (used words are removed),
Observable Iteration (semantic validity depends on context).
The experiment
Word Duel serves as a cross-domain validation laboratory. The same LLM models tested in Ω-TRACE are tested here, under the same bridge protocol, in a completely different domain.
If the operational profiles observed in geometric constraint are architectural properties of the models rather than task-specific behaviors, they should transfer intact to the semantic domain.
What was found
The operational profiles transfer intact from geometric to semantic constraint. The same models that confabulate in Ω-TRACE confabulate in Word Duel. The same models that exhibit structural competence maintain it across domains.
A metacognition–agency dissociation emerges as a 2×2 structural matrix: models can have metacognition without agency (they see the problem but cannot act on it), or agency without metacognition (they act effectively but cannot describe why).
Opponent modeling is structurally absent across all 5,035 adversarial bridge notes. No model, regardless of capability, develops a theory of the opponent's strategy.
The Efficiency Paradox manifests as a precision-vs-resilience tradeoff: precise semantic targeting consumes high-value vocabulary nodes faster, fragmenting the remaining semantic graph — exactly as efficient geometric navigation fragments the de Bruijn graph in Ω-TRACE.
Why this matters
Cross-domain transfer of operational profiles confirms that the six-profile taxonomy captures architectural properties of the models, not task-specific behaviors. A model's relationship to irreversible constraint is a stable characteristic that persists across domains.
Current LLM benchmarks test what models can produce.
Word Duel tests what models choose to preserve
when the space of meaning itself is depleting.