Glossary

The vocabulary the playbooks and research are written in.

AI Audit Trail: An AI audit trail is the chronological record of an AI system's inputs, retrieved context, and outputs, kept so a specific decision can be reconstructed after the fact. It is not the same as provenance, which tracks where a claim originated; the trail records how information was used. Compliance requirements bite at exactly this point: an auditor asks for the sources behind one decision on one day, and only a trail captured at the time can answer.
AI Compliance: AI compliance is being able to evidence, after the fact, how a system reached a given output and what it relied on. It is not a policy document or an approved-tools list. The requirement that actually bites is reconstruction: showing an auditor the sources behind a specific decision on a specific day, which is only possible if provenance was captured at the time rather than reconstructed later.
AI Memory: AI memory is context that survives across sessions and across tools, so an assistant does not start from zero each time or relearn what another tool already knows. It is not a longer context window, which extends a single session rather than persisting anything, and not chat history, which replays what was said rather than retaining what was established. Memory for one agent is also not memory for an organization; the two are different layers of the stack.
AI Provenance: AI provenance is knowing which source a specific claim came from, at the level of the claim rather than the answer. A cited answer tells you which documents were consulted; provenance tells you which sentence rests on which source. The difference matters the moment someone asks why a system said something, because an answer-level citation cannot be checked and a claim-level one can.
AI Verification: AI verification is checking what a model actually claimed against the sources it should have relied on, after generation. Grounding improves what goes in; verification tests what came out. The two are routinely conflated, and the distinction matters because a well-grounded system can still produce a fluent claim no source supports, and only a check after the fact will catch it.
ATOM: ATOM, the Augmented Transferable Operational Matrix, is Nucleus infrastructure that converts a government's regulatory frameworks from static documents into live, queryable intelligence that applications call at the point of decision. Compliance tools copy rules into software and go stale when the rules change; ATOM inverts the direction, so a regulator's update propagates automatically to every connected application, and any vendor building for a regulated market can plug in over API or MCP to ship compliant. Nucleus delivered ATOM's first client deployment in coordination with Rewaaj, integrating regulatory standards from roughly 15 countries.
COVE: COVE, or Chain of Verification, is a method that checks a generated answer claim by claim against its sources after the answer exists, rather than trying to retrieve better beforehand. Each claim is scored as source supported, unverifiable, or contradicted. Unverifiable and contradicted are deliberately separate states, because collapsing them either hides a hallucination or raises a false alarm. The method originates with Meta AI and ETH Zurich (arXiv:2309.11495) and is implemented and extended by Nucleus.
Compaction: Compaction is the lossy summarisation an AI system performs when its context window fills: it keeps what looks salient, silently drops the rest, and never reports which. It is the mechanism behind much of context rot. Research on long-context behaviour finds models retain only a small fraction of long inputs after compaction, which is why the failure produces no error signal, only answers that quietly stop honouring earlier constraints.
Context Engineering: Context engineering is the practice of deciding what a model sees before it answers: which sources, which slices of them, joined how, and in what order. Prompt engineering shapes the instruction; context engineering shapes the evidence. As systems move from single prompts to agents running unattended, it is the part that determines whether the output is grounded.
Context Layer: A context layer is the persistent, source-grounded organizational memory that sits between an organization's data and whichever model it uses, so every assistant and agent reads the same weighted, reconciled picture instead of re-deriving it per tool. It exists independently of any single agent or session. It is not a vector store, which retrieves passages without reconciling them, and not agent memory, which belongs to one agent and resets with it.
Context Poisoning: Context poisoning is the corruption of an AI system's persistent memory or retrieved context with false information, so the system keeps reasoning from it long after the attack. It is not prompt injection, which manipulates a single input; poisoning targets the ongoing knowledge an agent trusts. The defence is the discipline that makes context trustworthy at all: tiered trust, provenance on every claim, and unverified material never treated as fact.
Context Rot: Context rot is the gradual degradation of an AI system's working context over a long session, as earlier information is displaced, summarised, or silently dropped. The model does not report a gap; it fills it fluently, so the output stays confident while the grounding decays. It is not a context window size problem. Peer-reviewed research shows input length alone degrades performance even with perfect retrieval, so a larger window delays it rather than preventing it.
Contextual Intelligence: Contextual intelligence is what an AI system exhibits when it reasons over persistent, connected, trust-weighted organizational context rather than over whatever fits in a prompt. The premise is that dependable intelligence in deployed AI comes from governance of context, not from model scale: the model is a swappable reasoning engine, and the quality of its output is bounded by the context infrastructure underneath it.
Drift: Drift is an AI system's understanding diverging from the state of the thing it is reasoning about, so answers that were correct become quietly wrong as the underlying systems change. It is not the same as context rot. Rot degrades within a session as context is displaced; drift accumulates across sessions as the world moves and the system's picture does not.
Effective Context Length: Effective context length is the point in a model's input beyond which additional tokens stop improving output quality and start degrading it, and it is typically far shorter than the advertised context window. Benchmarks find only about half of models claiming 32K-token windows maintain performance at that length, and peer-reviewed work shows input length alone hurts performance even with perfect retrieval. The marketed window is a capacity claim; the effective length is what you can actually use.
Grounding: Grounding is supplying a model with the sources it should rely on before it answers, so the output is constrained by evidence rather than by the model's guesswork. It improves what goes into generation; it does not test what came out, which is verification's job. The two are parallel disciplines, not interchangeable ones: a well-grounded system can still produce a fluent claim no source supports.
Hallucination: A hallucination is a confident, fluent output that no source supports. It is usually described as a model defect, but the more useful description is a context gap: the model cannot perceive a missing piece as missing, so it experiences the gap as an ordinary generation problem and fills it. That is why better prompting does not fix it and why the fix has to sit outside the model.
Organizational AI Memory: Organizational AI memory is the persistent, permission-aware record of what an organization knows, decides, and corrects, held so any AI system can retrieve it at inference time. It is not a knowledge base, which stores content; memory also preserves who established what, when, and on what evidence. And it is not agent memory, which belongs to one agent and resets with it. Memory for one agent is not memory for an organization.
Regulatory Intelligence: Regulatory intelligence is a government's rules made machine-readable, queryable, and current, so compliance is checked at the point of decision rather than audited days later. It is not a compliance tool, which copies rules into software and goes stale when they change. The direction of the connection is the difference: the regulator's framework is the live source, and applications query it, so an updated standard propagates without anyone re-reading a PDF.
Sovereign AI: Sovereign AI is infrastructure that allows a government's regulatory frameworks, standards, and institutional knowledge to operate as live, enforceable context within any AI application in its jurisdiction. It is not defined by who builds the model or where the hardware sits; compute can come from any provider in any country. Sovereignty lives in the rules: when a nation's regulations are queryable infrastructure, reachable over API or MCP, every application operating under its authority can plug in and ship compliant.
Understanding Layer: The understanding layer is Nucleus's name for the context layer it builds: the layer between organizational data and foundation models where scattered sources become one coherent, verifiable account of what an organization knows. Storage holds the data and the model does the reasoning; the understanding layer makes the two compatible, adding memory, relationships across sources, grounding, and verification. It is an implementation of the context layer category, not a competing concept.
Working Layer: The working layer is everything an AI system uses during a task to produce output: the model's reasoning, agent orchestration, and each agent's session-to-session memory. It is the active inference region of the stack, and recent agent-memory progress lives entirely inside it. It is not the context layer, which persists above it independently of any single run. The failure modes differ: a weak working layer underperforms; a weak context layer produces outputs that are confidently wrong.