Concepts & glossary

Every term Kaizen uses, defined once. Pages elsewhere link back here rather than redefining a word in passing.

The core loop

Search — the non-generative surface. A query is retrieved and ranked, and you get source cards back. No language model writes the answer.

Chat — the generative surface. An assistant retrieves sources, reasons over them, and writes a cited answer, using tools along the way.

Turn — one exchange in a chat: your message, everything the assistant does in response, and the answer it commits. A turn is the unit of budget, policy, and audit.

Run — one execution of a turn by the runtime. A turn can be retried or recovered and still be the same logical run. A run produces exactly one terminal outcome.

Agent — a reusable configuration for how Kaizen should behave: a persona, a set of granted tools, a knowledge scope, and a mode. Every chat and every search is powered by one.

Mode — whether an agent belongs to the search or the chat surface. Agents are mode-scoped, so a Search agent never appears in the Chat picker.

Tool — something an agent can do rather than say: search, read a document, run code, create an artifact, manage a task, delegate to another agent.

Effort — how much reasoning budget a turn is allowed. Higher effort costs more time and tokens and is chosen per turn.

Knowledge

Corpus — your organization's collective indexed knowledge, drawn from connected sources. Visibility is governed by each source's own permissions.

Connector — a Kamiwaza-managed adapter to an external source (a docs site, a wiki, a tracker). Kaizen discovers connector instances at runtime rather than shipping a fixed catalog.

Document — a file you uploaded. Private to you, indexed separately from the corpus, and never mixed into it.

Chunk / passage — the unit of retrieval. Documents are split so a query can match the relevant part rather than the whole file.

Collection — a personal grouping of sources you can scope retrieval to.

Citation — a link from a claim in an answer back to the passage it came from. Clicking one opens the original.

Evidence authority vs navigation authority — a document result can be cited as fact. A chat or Project result can only help you find prior work; it never enters the citation ledger on its own.

Collaboration

Conversation — one chat, with one durable journal. Its identity never changes when it is shared, linked to a Project, or unlinked.

Journal — the append-only record of everything member-visible that happened in a conversation: messages, membership changes, artifacts, terminal outcomes. It is the source of truth for replay and reconnect.

Turns cache — a bounded convenience projection of the journal. Compaction can shorten the cache; it never edits the journal.

Project — a workspace that groups related chats and Library items by pointing at the existing items. A Project never copies an item, changes its owner, or replaces its access controls.

Direct share — named per-person access to one chat, independent of any Project.

Audience version — a monotonically increasing marker of a room's current membership. Every run carries the version it was authorized against.

Initiating participant / actor — the member whose identity, credentials, tools, and sandbox a given turn runs under. In a shared room this is per-turn, not per-room.

Presence — the short-lived "viewing" and "typing" indicators. Deliberately disposable: never journal content, never audit data.

Work that runs on its own

Task definition — mutable owner intent: what to run, with what inputs, on what schedule, with what limits.

Occurrence — one scheduled, manual, backfill, or test opportunity, including its exact calendar resolution and the dispatch decision made.

Run trace — the append-only, member-safe timeline of what a run actually did. Separate from the audit trail.

Lease — a short-lived, fenced claim that lets exactly one worker execute a run. Losing the lease removes side-effect authority.

Grant — a credential-free, revocable authorization fact bound to an exact owner, task, target, and run. Not a token; it cannot be replayed as one.

Trust

Identity envelope — the validated forwarded identity Kamiwaza hands Kaizen. The only identity source Kaizen accepts.

Guardrail — a backend-side check applied at defined points in a turn: budgets, citation discipline, redaction, loop control.

Write guard — the boundary an action with an external side effect must cross. An agent may propose; only the guard authorizes.

Capability ceiling — the maximum set of tools an agent may use, regardless of what a prompt asks for.

Fail closed — the rule Kaizen follows when authority cannot be proven: deny, rather than fall back to a weaker check or a broader identity.

Audit trail — the append-only record of who attempted or completed a consequential operation, its outcome, and a privacy-safe reason code.

Runtime

api-backend — the trusted service. Owns identity, permissions, retrieval, policy, budgets, and audit.

agent-runtime — the conversational loop. Powerful but secretless; it calls back into the backend for everything consequential.

Callback — a request from the runtime into the backend to actually execute something. The point where policy is applied.

Sandbox — the isolated, disposable environment where member code runs.

Compaction — always-on, system-managed shortening of long context so a conversation stays inside the model window.

Compression — the separate, opt-in, member-owned settings that shorten tool results or answer prose. Never the same thing as compaction.