# Enterprise Architecture Glossary

> Plain definitions of the terms in the atlas: control plane, static stability, fail-static, revocation window, Tier-0, KV cache, MCP and more.

- Web page: https://www.alesanalytics.com/architecture/glossary
- Markdown: https://www.alesanalytics.com/architecture/glossary.md
- Part of: [Enterprise Architecture Atlas](https://www.alesanalytics.com/architecture.md), section "Reference"
- Last verified: 13 October 2026
- Publisher: Ales Analytics

The terms the atlas uses, defined in a sentence or two each, with a link to the page that explains them properly. Where a term has a common meaning and a narrower one here, the definition gives the narrower one.

## A

**Admission webhook.** A service the Kubernetes API server calls before it accepts a write, to validate or change the object. It sits in the write path, so a webhook that is down with failurePolicy set to Fail blocks every write, including the fix. [Read more](https://www.alesanalytics.com/architecture/compute-and-kubernetes.md).

## B

**Blast radius.** Everything a failure or a compromise takes with it: the direct dependents, the transitive ones, and everything sharing the failed component's control plane, configuration, identity issuer or DNS, minus whatever is statically stable for long enough. [Read more](https://www.alesanalytics.com/architecture/failure.md).

**Blueprint (agent identity blueprint).** In Microsoft Entra Agent ID, the template and identity from which agent identities are created. It holds the credentials the agents authenticate with, and policy applied to it reaches every agent made from it, which makes it the blast-radius unit for agents. [Read more](https://www.alesanalytics.com/architecture/microsoft-365.md).

## C

**CAE (continuous access evaluation).** Microsoft Entra's event-driven token revocation. Tokens in CAE sessions can last up to 28 hours, and critical events such as a disabled account revoke them, but only at resources that take part. [Read more](https://www.alesanalytics.com/architecture/identity.md).

**Change data capture (CDC).** Reading a database's commit log to stream its changes to other systems. It preserves order per key or partition, not globally. [Read more](https://www.alesanalytics.com/architecture/data-and-storage.md).

**Confused deputy.** A service tricked into using its own authority on someone else's behalf, typically because a token was passed through instead of exchanged for one issued to the next audience. [Read more](https://www.alesanalytics.com/architecture/identity.md).

**Control plane.** The part of a system that creates, changes, authorizes and configures resources. It is usually shared and often global, and it is where the largest outages of 2024 and 2025 started. [Read more](https://www.alesanalytics.com/architecture/primitives.md).

## D

**Data plane.** The part of a system that does the work its resources exist for: serving requests, reading objects, forwarding packets. A well-designed data plane keeps working while its control plane is down. [Read more](https://www.alesanalytics.com/architecture/primitives.md).

**Decision 2.** In the two-decision model for AI over Microsoft 365, the separate decision that the AI may process an item the user is allowed to open, implemented by Purview DLP on the Copilot location. [Read more](https://www.alesanalytics.com/architecture/microsoft-365.md).

**DRA (dynamic resource allocation).** The Kubernetes mechanism for allocating devices such as GPUs by attribute rather than by name. Its core is stable; it does not enforce team quotas, so a queueing layer is still needed. [Read more](https://www.alesanalytics.com/architecture/compute-and-kubernetes.md).

## E

**E1 to E5 (equivalence classes).** The atlas's scale for how far one provider's service can stand in for another's: E1 equivalent, E2 equivalent with a different operating model, E3 structurally different, E4 unique to one provider, E5 the same word for different things. [Read more](https://www.alesanalytics.com/architecture/cloud-providers.md).

**Early and late binding.** Two ways to bind permissions to retrieval. Early binding stores ACLs in the index, which is fast and stale by the sync lag; late binding verifies each result at the source, which is exact and slower. [Read more](https://www.alesanalytics.com/architecture/ai-retrieval-and-agents.md).

**Error budget.** The amount of unreliability an SLO allows: one minus the SLO, multiplied by the window. At 99.9% over 30 days it is 43.2 minutes. [Read more](https://www.alesanalytics.com/architecture/numbers.md).

## F

**Facet delta.** A declared difference in one aspect (consistency, authorization, failure domain, lifecycle or namespace) between services that are otherwise equivalent. [Read more](https://www.alesanalytics.com/architecture/cloud-providers.md).

**Fail-static.** A failure mode in which a check keeps using its last known good decision or data, up to a declared maximum age, rather than failing open or closed. [Read more](https://www.alesanalytics.com/architecture/failure.md).

**Fencing token.** A number that increases every time leadership changes, attached to each write so that storage can reject writes from a leader that has been replaced. [Read more](https://www.alesanalytics.com/architecture/distributed-systems.md).

**FIC (federated identity credential).** A credential that lets a workload exchange a token from a trusted external issuer, such as a managed identity or a CI platform, for an Entra token, so no secret has to be stored. [Read more](https://www.alesanalytics.com/architecture/identity.md).

## I

**Idempotency key.** A value a caller attaches to a request so the receiver can recognise a repeat and do the work only once. It is what turns at-least-once delivery into an exactly-once effect. [Read more](https://www.alesanalytics.com/architecture/distributed-systems.md).

## K

**KV cache.** The keys and values a transformer stores for every token it has processed, reread for each new token. Its size grows with tokens and concurrent sequences, which is why serving language models is mostly a memory problem. [Read more](https://www.alesanalytics.com/architecture/ai-serving-and-gateways.md).

## L

**LKG (last known good).** The most recent configuration or decision known to work, which a system falls back to when a new one fails validation or cannot be fetched. [Read more](https://www.alesanalytics.com/architecture/failure.md).

## M

**MCP (Model Context Protocol).** The open protocol for giving AI models tools and context. Its 2026-07-28 revision removed protocol sessions, so state lives in server-minted handles and side-effecting tools need idempotency. [Read more](https://www.alesanalytics.com/architecture/ai-retrieval-and-agents.md).

**MRTR (multi round-trip request).** The MCP pattern in which a server returns an input_required result asking for more information, and the client retries the original request with the answers. [Read more](https://www.alesanalytics.com/architecture/ai-retrieval-and-agents.md).

## O

**OBO (on-behalf-of).** A token exchange in which a middle tier trades the token it received, issued to itself, for a new token issued to the next service, so the user's own token is never forwarded. [Read more](https://www.alesanalytics.com/architecture/identity.md).

## P

**PagedAttention.** A technique that manages the KV cache in fixed-size pages, as an operating system manages memory, removing fragmentation. Its authors reported 2 to 4 times the throughput at equal latency. [Read more](https://www.alesanalytics.com/architecture/ai-serving-and-gateways.md).

**PDP and PEP.** The policy decision point, which decides whether a request is allowed, and the policy enforcement point, which enforces that decision. In an AI system the model is never the PEP. [Read more](https://www.alesanalytics.com/architecture/identity.md).

**Principal access boundary (PAB).** A Google Cloud IAM policy that defines which resources a principal is eligible to access at all. It is checked before deny and allow policies, and fails closed. [Read more](https://www.alesanalytics.com/architecture/cloud-providers.md).

**Private endpoint.** A network interface with a private address that gives access to a platform service, made reachable by a private DNS zone answering the service's public name. Whoever owns the zone owns the path. [Read more](https://www.alesanalytics.com/architecture/networking.md).

## R

**RCD (Restricted Content Discovery).** A SharePoint site setting that keeps a site out of tenant-wide search and Microsoft 365 Copilot without changing who can open it. It replaces Restricted SharePoint Search, which retires on 31 January 2027. [Read more](https://www.alesanalytics.com/architecture/microsoft-365.md).

**RCP and SCP.** AWS Organizations policies that set permission ceilings. A resource control policy caps what can be done to resources; a service control policy caps what principals can do. Neither grants anything, and RCPs are evaluated first. [Read more](https://www.alesanalytics.com/architecture/cloud-providers.md).

**Regional compute, global state.** The anti-pattern in which a service runs regionally but reads data replicated globally, so one bad record can fail it in every region at once, as in Google's June 2025 incident. [Read more](https://www.alesanalytics.com/architecture/failure.md).

**Revocation window (W_revoke).** The longest time a removed permission can still be used: remaining token validity, plus decision-cache TTL, plus group-sync lag, plus index ACL-sync lag. It should be stated per data class. [Read more](https://www.alesanalytics.com/architecture/identity.md).

## S

**Static stability.** The property that a data plane keeps working, with no control-plane action, while the control plane is impaired. It usually costs pre-provisioned capacity: about 1.5 times peak across three zones. [Read more](https://www.alesanalytics.com/architecture/primitives.md).

**Storage and exposure classification.** Two classifications for derived data. Storage classification is always the highest of the inputs and governs where an index may live; exposure classification is per request, after trimming, and governs what an answer may show. [Read more](https://www.alesanalytics.com/architecture/ai-retrieval-and-agents.md).

## T

**Tier-0.** A component whose compromise or failure compromises or fails everything else: the identity provider and permission-granting identities, DNS, KMS and PKI, the edge, the AI gateway, CI/CD and IaC state, the backup control plane, and the time source. [Read more](https://www.alesanalytics.com/architecture/failure.md).

**TTFT (time to first token).** How long a language model takes to produce the first token of its answer. It is set mostly by prefill, which is compute-bound. [Read more](https://www.alesanalytics.com/architecture/ai-serving-and-gateways.md).

**Two-decision model.** The model for AI over enterprise content: Decision 1, can the user open the item, and Decision 2, may the AI process it. Both are needed, and they are enforced by different systems. [Read more](https://www.alesanalytics.com/architecture/microsoft-365.md).

**Type-1 decision.** A decision that is expensive or impossible to reverse, such as the choice of identity provider or tenant topology. It deserves a recorded exit cost and revisit trigger. [Read more](https://www.alesanalytics.com/architecture/principles.md).

## W

**Workload identity federation.** Letting a workload present a token from its own platform, such as a CI system or a Kubernetes cluster, and receive short-lived cloud credentials in return, so no static keys exist to leak. [Read more](https://www.alesanalytics.com/architecture/identity.md).

Should you have any questions regarding this page, please contact us at solutions@alesanalytics.com.

## Where to go next

- Previous: [Claim Verification Standard](https://www.alesanalytics.com/architecture/verification-standard.md)
- Overview and every page: [Enterprise Architecture Atlas](https://www.alesanalytics.com/architecture.md)
- The whole site in one file: https://www.alesanalytics.com/llms-full.txt