Grounding AI with Wittgenstein: From Language-Games to Epistemic Honesty (marcoeg.medium.com)

Researchers introduced LGDL (Language-Game Description Language), an open-source framework and MVP that reframes LLM hallucinations as a grounding problem inspired by Wittgenstein’s notion of language-games. Rather than pushing models toward universal, context-free fluency, LGDL lets developers encode bounded “games” (e.g., medical triage, contract review) with explicit participants, conversational moves, roles, and public criteria for correctness. The core design goal is epistemic honesty: models must signal uncertainty, negotiate meaning, and escalate when confidence thresholds aren’t met, instead of producing plausible-sounding but ungrounded answers. Technically, LGDL is a small domain-specific language compiled into three operational modes: (1) synthetic training-data generation to fine-tune models on rule-governed interaction patterns and calibrated confidence; (2) a deterministic interpreter/runtime contract that enforces domain rules, logs confidence, and blocks unsafe moves (e.g., no speculative cardiac diagnoses); and (3) runtime agents that combine the above with external capabilities for grounding. The implication for AI/ML is a shift from monolithic scale to bounded, expert-authored specialization: smaller or fine-tuned models can achieve safer, auditable behavior by participating in well-defined practices. The approach is speculative and requires domain expertise, but offers a concrete, philosophically grounded pattern for reducing silent hallucination in high-stakes applications.