ABSTRACT:Large language models (LLMs) generate fluent text but remain unreliable: they hallucinate facts, collapse under contradictions, and struggle to distinguish between uncertainty, falsity, and truth. To address this, we propose GRAIL (Grounded Reasoning and Inference Layer), a semantic framework that integrates an eight-state logic codex with ERA (Existence–Reasoning–Action), a structured pipeline for claim evaluation. ERA decomposes natural-language statements into three phases: ETQ (existence, time, question) for grounding and filtering, RCMR (reference, compare, memory, range) for semantic reasoning, and AE (action, evaluation) for downstream decision-making. GRAIL maps the outputs of this pipeline into eight logical states, extending beyond binary true/false to include “don’t know,” “not true,” “not false,” “contradictory,” and “both true and false.” This enables LLMs and embodied agents to handle uncertainty, paradox, and temporal reasoning in a structured manner. We outline a minimum viable implementation using parsers, knowledge graph lookups (e.g., Wikidata), numerical solvers, and factchecking APIs, and we discuss applications in hallucination filtering, embodied robotics, and decision support. GRAIL is not a step toward AGI but toward reliable, transparent, and human-compatible AI systems capable of operating in complex environments with bounded trust.