Definition
Context Hallucination
A hallucination that occurs not from false information, but from failure to account for relationships between documents, where one document modifies, contradicts, or conditionally applies to another.
From the episode “Legal AI: Why Lawyers Are Finally Free to Think”Hear it on the podcast1 episode
Learn the language of AI-native founders
Every week: the new terms, tools, and workflows founders are using, straight from the podcast.
Related terms in AI Fundamentals
Constitutional LayerThe rules, values, and ethical principles baked into each agent at instantiation. The foundational constraint set that shapes all downstream behavior before any prompt is processed.Context ArchetypeA named configuration of rules governing how an AI generation system behaves in a specific mode — for example, 'creative exploration' vs. 'compliance review.' The same underlying model produces reliably different outputs depending on which archetype is active. In Knapsack's Intelligent Product Engine, archetypes are one of three components (alongside real-time artifact monitoring and a rules engine) that make AI generation predictable and enterprise-appropriate rather than open-ended and unpredictable.Context engineeringStructuring the inputs and memory an LLM receives. Ross Barnes's position is that the difference between reliable and hallucinatory output is almost always context, not model choice.Context GroundingThe practice of providing an AI coding agent with accurate documentation of the existing codebase before asking it to generate or modify code. Without grounding, agents operate on generic code patterns and produce architecturally inconsistent output that may be technically correct in isolation but incompatible with the existing system. BrainGrid generates five grounding documents from the actual repository — architecture overview, key workflows, data structures, directory structure, API patterns — so the agent extends what's there rather than inventing new patterns.Context windowHow much input, such as text, documents, or conversation history, an AI model can take into account at one time.Contrastive learningA training technique where a model learns by comparing similar and dissimilar examples, figuring out what makes things alike or different.Data augmentationExpanding a training dataset by creating modified copies of existing data, like flipped images or reworded sentences, instead of collecting new data.Data Collection (Local AI)A curated subset of a user's files — PDFs, documents, images, voice memos — that is analyzed once by a local AI system and made queryable without any cloud access. The unit of organization for privacy-first AI interaction with personal or proprietary data.