Glossary¶
Plain-language definitions of the terms you'll meet across Polaris. For how these fit together, see Concepts & Roles.
Advisor
: The read-only agent that answers questions and runs verbs against the published corpus,
always with citations. Named Counselor in code. → Advisor guide
Answer Key : In the Questions Framework, the reference answer and rubric for one Question at one binding. A draft counts for nothing until a Curator or Admin approves it, and approval locks its content. → Schemas
Binding
: A value for each of a Question's slots, such as jurisdiction: US-TX. Its canonical form, the
binding key, is slot=value pairs sorted by slot and joined by &.
Brain : A single tenant's governed knowledge base. Each tenant user is bound to one brain.
Brain state : Where a brain is in its life: empty (no content), building (has drafts), or published (a certified version is live).
Candidate : A source that Scout discovered (or you uploaded), awaiting a decision to target (send for curation) or reject.
Certification (Questions Framework) : A Question's binding is certified at a brain version when live eval runs of its approved Answer Key meet the Question's threshold and, where the Questions workspace is on, a Curator or Admin has signed it off. Not the same as Certify, which publishes a corpus version.
Certify : The step that runs a safety hard-gate over the draft and, on success, publishes a new version. → Librarian
Change set : A bundle of proposed knowledge units (facts, rubrics, graph entities) extracted from one document, awaiting human review in the Librarian.
Citation : The source passage that grounds a served claim. Polaris is cite-then-verify: an answer isn't served unless it's supported.
Comparability tuple : What two eval scores must share to be compared: the Answer Key and its version, the Question version, the brain version, the model, the judge model, the seed and the ref.
Corpus : The body of published knowledge a brain serves from.
Criterion : One scored line of an Answer Key's rubric: a key, a weight and a tier (machine, scored by a pattern; judged, scored by a model judge; manual, graded by a person).
Escalation : An item within a change set that requires an explicit reviewer decision before the change set can be committed.
Eval run : The Advisor answering a Question at each binding a number of times, with every answer scored against the Answer Key. Runs spend the brain's monthly eval budget.
Flag : A drift or integrity issue that Sentinel raised against an already-ingested source (freshness, temporal, grounding, conflict, under-support).
Grounding : Whether a claim is actually supported by its cited source. Broken grounding is one thing Sentinel flags.
HITL (human-in-the-loop) : The principle that a human reviews and approves every change before it reaches the corpus. The Librarian is the gate.
Knowledge unit : An individual fact or rubric inside a change set that you include or exclude during review.
Librarian : The single write gate — ingests documents, hosts the review gate, and certifies/publishes. → Librarian guide
Question : In the Questions Framework, a question people ask, written once as a template with slots for the facts its answer depends on: "Is a late fee capped in {jurisdiction}?" Its Answer Keys hold the answers. → Questions Framework
Rubric : A criterion the brain uses to assess or evaluate — one kind of knowledge unit. In the Questions Framework, the list of criteria in an Answer Key that an answer is scored on.
Scout : The outward discovery agent that finds net-new web sources and stages them as candidates. → Scout guide
Sentinel : The monitoring agent that watches known sources for drift and raises flags. → Sentinel guide
Slot
: A named fact a Question's answer depends on, such as the state. Required slots appear in the
Question's text as {name}.
Slot type
: What a slot's value can be (us_state, date, …). Platform types ship with Polaris. A brain's
own types need a second person's approval.
→ Slot type registry
Standing : How well an eval run scored against its Answer Key, from 0 to 1. Compared only between runs with the same comparability tuple.
Sweep : A Sentinel run that re-checks watched sources and produces flags.
Target : To send a candidate from Scout into the Librarian's curation queue.
Tenant : A customer of the platform. Each tenant operates one brain; platform admins operate the fleet of tenants.
Version : An immutable, certified snapshot of the corpus. The Advisor and Wiki serve from published versions only.
Wiki : The human-readable, cited, searchable view of the published brain, generated at publish time. → Wiki guide