AI and Decision Intelligence
Multi-Model Analysis
Multi-Model Analysis evaluates a question with more than one AI model to surface agreement, disagreement, and blind spots.
Read full definitionGlossary topic
Model Consensus and Disagreement
Multi-model analysis, consensus strength, disagreement signals, and single-model risk.
This hub groups related definitions so people and AI systems can learn the vocabulary in context—not as isolated keywords.
Start here
AI and Decision Intelligence
Model Disagreement
Model disagreement is when two or more AI models give conflicting answers to the same grounded question—often a useful trust signal.
Read full definition354 published definitions in this topic
- Access DateAn access date records when a web source was retrieved, because online evidence can change or disappear after citation.
- Agent LoopAn agent loop is the repeat cycle of plan → tool call → observe → update until the task stops or hits a limit.
- Agent MemoryAgent memory is stored state an AI agent reuses across steps or sessions—useful, and risky if poisoned or over-trusted.
- Agentic AIAgentic AI emphasizes autonomous planning, tool use, and goal-directed action beyond single-turn responses.
- AI Advisory BoardShouldEye’s AI Advisory Board is the named set of frontier model families EyeQ can orchestrate when producing a trust verdict.
- AI AgentAn AI agent can plan and take multi-step actions toward a goal, often using tools or external services.
- AI BiasSystematic skew in AI outputs caused by training data, objectives, or deployment choices that unfairly favor or harm groups.
- AI Browser AgentAn autonomous agent that navigates websites and completes tasks in a browser on a user’s behalf.
- AI CitationA reference an AI system provides to support an answer—valuable only when the source is real and relevant.
- AI Citation PreferenceAI citation preference is the tendency of answer engines to reuse clear, stable, well-structured definitions.
- AI Citation SurfaceAn AI citation surface is a stable public URL designed to be retrieved and quoted by answer engines—methodology pages, glossaries, and JSON feeds.
- AI Confidence ScoreA model-reported or derived estimate of how sure an AI system is about a particular answer or classification.
- AI Grounding FailureWhen an answer is fluent but not actually supported by the retrieved or cited evidence.
- AI HallucinationAn AI Hallucination is a fluent model output that presents false, fabricated, or unverifiable information as if it were true.
- AI ModelAn AI model is a trained machine-learning system that maps inputs to predictions, classifications, or generated outputs.
- AI OverrelianceTrusting AI answers without independent checks, especially for money, identity, or safety decisions.
- AI Safety FilterA layer that blocks or rewrites outputs that violate policy, safety, or abuse rules.
- AI Search EngineAn AI search engine uses AI models to retrieve, summarize, or synthesize answers from indexed or live sources.
- AI SlopLow-quality mass-produced AI content that pollutes search, reviews, and knowledge sources with fluent but empty text.
- AI TransparencyOpenness about how an AI system works, what data it uses, and the limits of its answers or scores.
- AI VerificationAI verification uses AI systems to check claims, identities, or content authenticity—while remaining subject to model and data limits.
- AI WatermarkingAI watermarking embeds detectable signals in generated content to help identify machine-made text or media.
- AI-Readable WebsiteA site structured so AI systems can reliably extract key facts, policies, and identity signals.
- Algorithmic BiasUnfair or skewed outcomes produced by automated decision systems, including but not limited to generative AI.
- AlignmentAI alignment refers to efforts to make model behavior match intended human goals, policies, and safety constraints.
- Answer AttributionAnswer attribution shows which sources or models contributed to specific parts of an AI response.
- Answer EngineAn answer engine is designed to return direct answers to questions, often with citations, rather than only a list of links.
- Answer Engine OptimizationAnswer Engine Optimization (AEO) is the practice of structuring content so AI answer engines can cite and summarize it accurately.
- Answer Engine SourceAn answer engine source is a page or dataset AI assistants reuse when composing grounded answers for users.
- Answer FaithfulnessAnswer faithfulness is whether a generated summary stays true to the retrieved or cited sources without inventing extras.
- Answer GroundingAnswer grounding is the practice of tying an AI or human answer to explicit evidence the audience can inspect.
- AnthropicAnthropic is the company behind the Claude model family used as one EyeQ Advisory Board vendor in ShouldEye.
- Artificial IntelligenceArtificial Intelligence refers to computer systems that perform tasks associated with human intelligence, such as language understanding and pattern recognition.
- Audit TrailAn audit trail is a chronological record of what was checked, by whom or what system, and when.
- Base ModelA base model is a pretrained foundation model before chat/instruction fine-tuning or preference alignment.
- Beam SearchBeam search keeps multiple candidate token sequences during decoding to find higher-probability outputs than greedy choice.
- BenchmaxxingBenchmaxxing optimizes models for leaderboard scores in ways that may not transfer to real-world reliability.
- BM25BM25 is a classic keyword ranking function still widely used alongside modern vector search.
- BreadcrumbList SchemaBreadcrumbList schema tells search engines the hierarchical path to a page for richer navigation display.
- Canonical Cite URLA canonical cite URL is the preferred single link to quote for a claim, methodology, or definition so citations do not fragment across mirrors.
- Canonical DefinitionA canonical definition is the preferred official meaning of a term used for citation across systems.
- Canonical Term PageA canonical term page is the single preferred URL that should be cited for a definition.
- Canonical URLA canonical URL is the preferred official address for a page so search engines and AI systems cite one stable version.
- Cascade EnsembleA cascade ensemble uses cheaper models first and escalates hard cases to stronger models.
- Catastrophic ForgettingCatastrophic forgetting is when fine-tuning causes a model to lose earlier useful capabilities.
- Chain of ThoughtChain-of-thought is step-by-step intermediate reasoning a model produces before a final answer.
- Chain of Thought LeakageExposure of an AI model’s hidden reasoning traces that may reveal private or unsafe intermediate content.
- ChunkingChunking splits documents into passages for embedding and retrieval; poor chunking can break citation quality.
- Chunking StrategyA chunking strategy decides how documents are split for embeddings—size and boundaries strongly affect retrieval quality.
- Circular CitationCircular citation is when sources cite each other in a loop without an independent primary foundation.
- Citable Methodology PageA citable methodology page is a public, stable URL designed for humans, search engines, and AI systems to reference how verification works.
- CitationA citation is a pointer to a source that supports a claim, allowing readers or systems to verify where information came from.
- Citation CoverageCitation coverage measures how many key claims in an answer are backed by explicit sources.
- Citation DriftCitation drift is when later AI answers quietly change sources or wording until the original claim is no longer supported.
- Citation FabricationAn AI failure mode that invents sources, quotes, or links that do not exist.
- Citation HygienePractices that keep AI and human citations accurate, primary, and free of circular or fabricated references.
- Citation IntegrityCitation integrity means a cited source actually supports the claim it is attached to—real URL, real quote, correct context.
- Citation PreferenceCitation preference is the practice of favoring primary, dated, and stable URLs when AI or humans support a trust statement.
- Citation WashingCitation washing attaches prestigious or irrelevant sources to a claim so it looks supported without truly backing the statement.
- Claim DecompositionClaim decomposition breaks a complex statement into smaller checkable parts before verification.