Glossary

The vocabulary, defined once.

The published glossary renders from the same source as the product, so a term on this page means exactly what it means on the screen you will see later. Every term has its own page.

Hierarchy Governance Evidence Runtime Catalogue Engineering Commercial 30 terms

The four levels

Agent: The smallest unit of work in KrewOS. One Agent has one job, a typed input and output contract, a declared model profile, an explicit set of Tools it may call and an explicit set of effects it may cause. Agents do not decide their own scope.
krewos.ai/glossary/agent
Pod: A coordinated team of Agents that produces one composite result — for example an Evidence Research Pod of four Agents running eleven retrievals and reconciling them. Pods nest, collapse in the canvas, and carry their own typed contract to the rest of the crew.
krewos.ai/glossary/pod
Crew: A deployed, running instance of a Crew Template inside one Project and Environment, bound to a specific immutable Release. "The crew" is what your operators talk about; the template is what your builders edit.
krewos.ai/glossary/crew
Crew Template: A reusable, versioned, end-to-end workflow — Agents, Pods, routing, human wait states, declared effects, evidence floor and evaluation suite — designed to be configured rather than rebuilt. Six ship as production templates; more come with Packs.
krewos.ai/glossary/crew-template
KrewOS Pack: A licensed vertical package: Crew Templates, Knowledge Packs, deterministic policy packs, evaluation datasets and dashboards for one industry and jurisdiction, maintained and dated by its publisher. Licensed annually and separately from the platform.
krewos.ai/glossary/krewos-pack
Crew Manifest: The canonical, versioned, machine-readable definition of a crew. The visual canvas is an editor over the manifest, plain-English changes compile into it, and Git commits resolve to it. It can be validated, diffed, exported, tested and deployed. The schema is published as a public contract.
krewos.ai/glossary/crew-manifest

Evidence and abstention

Evidence: The set of retrieved sources supporting a specific claim, each carried with its authority tier, its date and its retrieval context. Evidence travels with the output all the way to the reviewer and into the audit export — it is not a debug artefact.
krewos.ai/glossary/evidence
Sufficiency Score: A score assigned to an evidence set against the specific claim it is asked to support, combining source authority, recency, coverage and agreement. It is compared against the crew's declared evidence floor at run time to decide whether to answer or abstain.
krewos.ai/glossary/sufficiency-score
Evidence floor: The minimum sufficiency score a Crew Template declares for its claims. Below the floor the crew must not answer. The floor is a manifest property, versioned and diffable, not a prompt instruction — so it cannot be argued away by the input.
krewos.ai/glossary/evidence-floor
Abstention: The structured result returned when retrieved evidence falls below the declared floor. It names what could not be established and which sources would resolve it, and routes to a human knowledge-gap queue. Abstentions are reported separately from failures, because a missing source is not a broken crew.
krewos.ai/glossary/abstention
RAG: Retrieval-augmented generation: grounding model output in retrieved source content rather than parametric recall. In KrewOS, retrieval is always attributable — every retrieved chunk keeps its source, authority and date so the claim it supports can be traced.
krewos.ai/glossary/rag
Knowledge Pack: A governed collection of source documents indexed for retrieval, with access control lists, freshness windows and dated currency assertions. Knowledge is not Memory: Knowledge is source-backed organisational content; Memory is task state across a run.
krewos.ai/glossary/knowledge-pack

Control, effects and autonomy

Action: A step that causes an externally visible side effect — sending, filing, posting, paying, publishing. Every Action requires an idempotency key and a declared compensation behaviour, and may require human approval. Distinct from a Tool, which only reads or computes.
krewos.ai/glossary/action
Effect Declaration: The statement each component makes about what it reads, writes, spends and causes externally. The compiler statically proves those properties before a Release exists — including that no path runs from untrusted input to an external side effect without an intervening approval gate. If the property does not hold, there is no release to deploy.
krewos.ai/glossary/effect-declaration
Progressive Autonomy: The mechanism by which a crew earns a lower review rate against measured accuracy at a published confidence bound, and loses it automatically on breach. Four stages: full review, supervised sampling, extended sampling, and a regulated floor that never graduates.
krewos.ai/glossary/progressive-autonomy
Review Sampling: Risk-weighted selection of outputs for human review once a crew has graduated past full review. Never uniform — high-risk, high-value and low-confidence outputs are sampled far more heavily. Reviewer behaviour is itself measured, because a sampling regime is only as sound as the reviews calibrating it.
krewos.ai/glossary/review-sampling
Correction Record: The immutable record of a human approval, rejection or edit, tied to a named reviewer, a run, a release and a component version. Append-only: corrections supersede, they never rewrite. Expert-confirmed corrections are promoted into the crew's evaluation suite, which is why the suite compounds.
krewos.ai/glossary/correction-record
Evaluation Suite: The versioned set of cases and scorers a crew must pass before a Release can be promoted. Suites start from the Pack's datasets and grow out of real disputed cases from your own work, which is why a competitor cannot copy yours.
krewos.ai/glossary/evaluation-suite

Platform and deployment

Control Plane: The central layer holding identity, the catalogue, manifests, policies, evaluations, releases and audit. It stays under one governance model regardless of where execution happens. Availability is committed at 99.9% monthly where the contract carries the commitment, and a month below it is credited automatically.
krewos.ai/glossary/control-plane
Execution Plane: The layer that runs the compiled plan and holds run state. It can sit in KrewOS Cloud, a dedicated environment, your own VPC, a hybrid topology or your own infrastructure. Splitting the planes is what lets you keep one governance model and still satisfy residency.
krewos.ai/glossary/execution-plane
BYOK: Bring your own key. Register your model provider credentials as encrypted secret references so inference runs against your account at your negotiated rates and your provider bills you directly. Platform credits are unchanged; the model meter reads zero. Keys never appear in prompts, logs, manifests, exports or support tools.
krewos.ai/glossary/byok
MCP: Model Context Protocol. KrewOS speaks it in both directions: deployed crews are callable as standard tools by external assistants with policy, evidence and approval enforced by the runtime rather than the caller, and approved external MCP servers are consumable as Tools under the same registry, permission and audit controls.
krewos.ai/glossary/mcp
Linked Project: A project exported to Git that continues to round-trip with the platform. Changes made in the repository return through a reviewed diff and an approval workflow before they touch a manifest, so the engineering workflow does not bypass governance.
krewos.ai/glossary/linked-project
Ejected Project: A project exported to Git and detached permanently. The repository — manifest, components, schemas, prompts, policies, evaluations, tests, deployment files — is yours outright, and the runtime APIs still work. Eject rights exist so the lock-in question has a real answer.
krewos.ai/glossary/ejected-project
Vibe Coding: Describing an intended change in plain English and letting the platform produce a change plan, a visual diff, a validation result and a test run for you to accept, amend or reject. It never edits production directly. Buyer-facing pages say "describe the change in plain English"; engineers call it Vibe Mode.
krewos.ai/glossary/vibe-coding

Assurance and commercial

Golden Crew: An internal reference crew held at a known-good state and used to validate platform releases before they reach tenants. It is a validation instrument, not a customer-facing concept — on customer surfaces the equivalent idea is simply a reference crew.
krewos.ai/glossary/golden-crew
Maintained Pack: A KrewOS Pack whose publisher commits to a stated update cadence, dated currency assertions on its knowledge, and change notifications to entitled tenants when a source materially changes. The alternative is a Pack that was accurate on the day it was written.
krewos.ai/glossary/maintained-pack
Outcome Unit: The unit of business outcome a crew declares — a matter reviewed, a reconciliation closed, an article published, an advisory issued. Cost per outcome and human minutes displaced are reported against it, from the same records that enforce your budgets.
krewos.ai/glossary/outcome-unit
Credit: One unit of metered platform consumption, prepaid and drawn down as crews run. One credit is one Singapore cent. Two meters draw on one wallet: platform credits, always charged, and model credits, which read zero under BYOK.
krewos.ai/glossary/credit
Availability service credit: The remedy owed when a month falls below a committed availability figure. Despite the name it is money against your next invoice, not platform credits: it creates no lot, never expires, and is paid out in cash if you leave before it is used up. Issued automatically from the published attainment, on the published schedule.
krewos.ai/glossary/availability-service-credit

Every entry, on its own page

Agent Pod Crew Crew Template KrewOS Pack Crew Manifest Evidence Sufficiency Score Evidence floor Abstention RAG Knowledge Pack Action Effect Declaration Progressive Autonomy Review Sampling Correction Record Evaluation Suite Control Plane Execution Plane BYOK MCP Linked Project Ejected Project Vibe Coding Golden Crew Maintained Pack Outcome Unit Credit Availability service credit Terms we will not use Terms we are careful with Where the vocabulary comes from

Terms we will not use

"Fully autonomous", "AI employee", "no hallucinations", "MAS certified", "PDPA compliant" as a property of software. Each is either unsupportable or contradicts a mechanic the product actually enforces. We say abstains rather than speculates, PDPA-aligned, and aligned with named MAS guidance.

Terms we are careful with

KrewOS Hub is a catalogue, not a marketplace. A prepaid purchase of credits is a bundle or top-up, never a "credit pack" — Pack is a reserved product object. Tools read; Actions cause effects; the two are never collapsed into "integrations".

Where the vocabulary comes from

One source, rendered into the product UI, the documentation and this page together. Vocabulary drift between a marketing page and the screen a user later sees is one of the cheapest trust signals to get right and one of the most common to get wrong.