Glossary

Plain-language definitions of the terms used across the Kinetic Gain pillar guides, spanning identity, data governance, AI security, FinOps, and signed provenance. Each term links to the guide that goes deep.

Identity & Access Data Governance AI Security FinOps Provenance / KGP

Identity & Access IAM

RBAC role-based access control: An access model that grants permissions to roles, and assigns people to roles, rather than granting permissions to individuals directly. It keeps entitlements manageable at scale by making the role the unit of access.
ABAC attribute-based access control: An access model that decides permissions from attributes of the user, resource, and context (department, sensitivity, time, location) evaluated against a policy, rather than from static role membership. More expressive than RBAC, and harder to reason about.
SCIM System for Cross-domain Identity Management: A standard protocol for automatically provisioning and deprovisioning user accounts across applications, so an identity created or removed in one system propagates to the others. It is the plumbing that makes joiner and leaver automation real.
Joiner-Mover-Leaver (JML) identity lifecycle: The three lifecycle events that change a person's access: joining (grant birthright access), moving (change entitlements to match a new role), and leaving (revoke everything). Most access risk lives in the mover and leaver events that never fully complete.
Birthright access: The baseline set of access a person receives automatically on joining, based on attributes like department and job function, before any specific request. Getting birthright right reduces the volume of one-off access requests.
Least privilege: The principle that a person or process should hold only the access it needs to do its job, and no more. It shrinks the blast radius of a compromised account or a hijacked agent, which is why it is a core defense against many attacks.
PAM privileged access management: The controls around high-power accounts (admins, root, service accounts): vaulting their credentials, granting elevated access just in time, and recording what was done with it. Privileged accounts are the highest-value target, so they get the tightest controls.
Deprovisioning: Removing a person's access when they leave or change roles, across every system that granted it. Incomplete deprovisioning leaves orphaned access that no longer maps to a current need, a standing risk auditors look for.
Separation of duties SoD: A control that splits a sensitive process across two or more people so no single person can complete it alone, for example the requester of access cannot also approve it. It prevents both error and self-dealing.

Data Governance DATA

Data lineage: The map of how each field flows from its source, through every transformation, to the dashboards, models, and exports that use it. Column-level lineage gives an exact blast radius for a change and traces a person's data to every downstream copy.
Data contract: An enforceable agreement about the shape, meaning, and guarantees of a dataset between the team that produces it and the teams that consume it. It turns silent schema drift into a checkable, versioned promise.
PII redaction: Automatically detecting and removing personally identifiable information from data before it reaches a place it should not be, such as a log, an analytics store, or a public page. It reduces the surface a breach or a subject request has to cover.
Retention schedule: A policy that maps each data category to how long it is kept and on what lawful basis, defaulting to delete when no longer needed. Data held past its usefulness is a liability, not an asset, because attackers exfiltrate what exists.
Soft delete vs hard delete: Soft delete flags a record as removed while the data stays present and recoverable. Hard delete physically removes it so it cannot be recovered. An erasure obligation contemplates hard delete, propagated to every copy.
Crypto-shredding: Rendering data unrecoverable by destroying the encryption key rather than the ciphertext. It is the standard way to satisfy an erasure obligation for data sitting in immutable, append-only backups that cannot be edited in place.
Legal hold: A directive to preserve data relevant to litigation, investigation, or audit, which suspends normal deletion for the affected records until it is lifted. It overrides the retention schedule, because deleting data under hold risks spoliation.

AI Security AI-SEC

Prompt injection: An attack where adversary-controlled text is read by a language model as instructions to follow rather than data to process. Direct injection comes from the user's input, indirect injection rides in on content the model later reads, such as a web page or email.
Confused deputy: The pattern where an agent holds credentials and tool access the attacker does not, so an injected instruction executes with the agent's authority rather than the attacker's. The privilege gap is what turns a text trick into real impact.
RAG retrieval-augmented generation: An architecture that retrieves relevant documents at query time and feeds them to a language model as context, so answers are grounded in a specific corpus. It introduces a supply of untrusted content into the model's context, which is its own attack surface.
AI-BOM AI/ML bill of materials: A machine-readable inventory of everything that went into an AI system: which model and version, its training or fine-tuning data, and its software dependencies. It extends the software bill of materials (SBOM) idea into the ML stack.
Backdoored weights: Model weights trained or fine-tuned so the model behaves normally in general but produces attacker-chosen outputs on a specific trigger. Because the behavior hides in opaque tensors, it survives redistribution and is invisible in a file listing.
Safe serialization safetensors: A model file format that stores only tensor data and cannot execute code when loaded, unlike pickle-based formats that deserialize into live objects. Preferring formats that cannot run on load removes a whole class of malicious-model-file risk.
Model card: A document that records a model's intended use, training data, and known limitations, giving its lineage a human-readable home. It is part of establishing provenance for a model you did not train yourself.

FinOps FINOPS

Cost allocation: Attributing shared infrastructure spend back to the teams, apps, and cost centers that consumed it. On a shared Kubernetes cluster this is hard, because nodes are billed whole but shared by many teams' pods.
Showback vs chargeback: Showback attributes cost and makes it visible to each owner without moving money, a mirror. Chargeback bills the cost back to a team's budget, adding real financial accountability. Most organizations start with showback and graduate to chargeback once the data is trusted.
Requests vs usage: In Kubernetes, requests are the capacity a pod reserves and holds out of the shared pool, while usage is what it actually consumes. Allocating cost on usage alone undercharges over-requesters, so many teams allocate on the maximum of request and usage.
Rightsizing: Adjusting the resources allocated to a workload to match what it actually needs, reclaiming reserved-but-unused capacity. It is the main lever teams pull to reduce their allocated cost once allocation makes the waste visible.

Provenance / Kinetic Gain Protocol KGP

Evidence object: A machine-readable record that attaches provenance to a claim: the claim text, its source, and a verification block with a content hash and optional signature. It is the core artifact a KGP-conformant page emits.
Content hash: A reproducible sha256 of the exact claim text, so any consumer can recompute it and confirm the claim was not altered. It is what makes a claim tamper-evident without requiring a signature.
Conformance level: One of Attach (1), Verify (2), or Sign (3), describing how strongly a page's provenance can be checked. You pick the lowest level that carries the weight the claim needs.
ed25519 signing: A modern public-key signature scheme used at KGP Level 3 to sign a claim's canonical form, with the public key published under the site's /.well-known/. A valid signature proves which entity stands behind a claim, not that the claim is true.
JCS JSON Canonicalization Scheme, RFC 8785: A deterministic way to serialize JSON (sorted keys, no insignificant whitespace) so the same data always produces the same bytes. It is what a signature is computed over, so both signer and verifier reconstruct identical input.