LLM

LLM refers to neural network models trained on massive text corpora to predict and generate natural language. Examples include GPT, Claude, Gemini, and Llama. LLMs can perform a wide range of language tasks—summarization, translation, question-answering, code generation—by learning statistical patterns from training data.

For Coscientist, LLMs are the engine that performs contemplation labor: proposing hypotheses, gathering evidence, finding counterexamples, and structuring arguments. Because LLMs can read any language, they enable cross-linguistic synthesis as a native capability.

However, LLMs have fundamental limitations. They optimize for plausible next tokens, not for truth. They can hallucinate: producing confident, coherent text that is factually wrong. They are susceptible to the fluency trap: smooth prose that masks errors. They share training data, so agreement among models may reflect correlated bias rather than independent verification.

This is why Coscientist treats LLMs as tools, not oracles. The Operator retains sovereignty; the epistemic protocol layer enforces traceability and rebuttal-first search; and the Multi-AI Consensus Protocol uses model disagreement as a signal for closer inspection. LLMs do the search and structuring; humans do the verification and decision.