knowledgev8 speaks OKF — the Open Knowledge Format — today, and is adopting the v0.2 trust-signal vocabulary: machine-readable answers to "who wrote this, who checked it, and is it still true?" The roadmap also tracks the operational company-brain surfaces around Workspace, Connectors, Workforce, and Review. Near-term public work is summarized below; detailed sprint planning is kept private.
The trust foundation the beta runs on — none of this is aspirational.
Every concept moves through draft → reviewed → authoritative → deprecated. Deprecated concepts drop out of retrieval and the graph, but stay preserved in export history.
statusAgent-generated knowledge always lands as a draft. Only a person can promote it — generated is never the same thing as verified.
generated ≠ verifiedEvery curated concept records the document or URL it was distilled from, with connector-level provenance across all nine connectors.
Source changes are tracked and affected concepts land in a re-curation queue, so drift gets caught instead of quietly served.
Lossless export/import round-trip of your whole knowledge base as OKF bundles — unknown fields are preserved, not dropped.
okf_version: 0.2Exported bundles carry the v0.2 trust-signal vocabulary — generated, verified, sources, and status — so trust travels with the knowledge.
okf_version: 0.2Retrieval is scoped to what the asking human can see, PII masking keeps sensitive values out of the brain, and every access is audit-logged.
See exactly where each concept came from — the source document, connector, and distillation trail — right on the concept page.
Who promoted a concept and when, surfaced next to the concept itself instead of buried in the audit trail.
verified {by, at}Tell us what's wrong (or right) without leaving the product — beta feedback routes straight to the team.
Ask for answers drawn only from human-reviewed knowledge, or widen to machine-confirmed and unverified when speed beats certainty.
min_trustConcepts carry a date after which they need a fresh human look — set per source as a content shelf life, stamped at promotion, or imported — with an optional archive mode that keeps past-due knowledge out of answers until it's re-verified.
stale_afterNear-duplicate detection at ingest and across the corpus, human-resolved supersede with lineage links, and contention-collapse so retrieval serves the freshest verified copy.
Workspace admins can define which input and output modules are available, see active source counts, review setup requirements, and jump straight into source creation.
connectorsSample AI employees now open real configuration flows, with detail pages showing purpose, required modules, watched sources, related work, and linked standing agents.
workforceReview brings drafts, standing-agent drafts, re-verification work, pending skills, duplicate pairs, answer-quality signals, source-quality findings, and connector setup gaps into one operational queue.
reviewEvery sample AI employee has render coverage, standing-agent route validation, and durable template-id persistence so template launch remains governed by normal standing-agent rules.
KVB-009Review's answer-quality queue follows the same admin-only exposure rule as Insights, with route coverage proving non-admin pages do not render weak, no-answer, or rated-down question text.
KVB-010Workspace, Workforce, and Review empty/setup-gap states link back to Connectors, and where the missing module is known they deep-link to that connector card.
KVB-011Sources now have compact scan rows, an expandable Edit panel, connector-specific configuration edits, write-only credential rotation, sync cadence, auto-curate, and shelf-life controls.
workspaceOrg rules and security practices served to coding agents at handshake (MCP instructions + API), human-verified content only.
onboardingDocument- and chunk-level security ACL filtering inherited from enterprise sources and enforced at query time so agents never see unauthorized knowledge.
security_aclsBidirectional Action Gateway attestation recording cryptographic execution proofs and policy citations for AI employee compliance.
attestationConcepts distilled from several documents cite each of them, with footnote-style citation mechanics in answers.
sourcesNear-term hardening: strengthen Workforce template coverage, finish sensitive-text gating, link connector empty states, run DB-backed validation, and keep expanding the OKF citation model.
Detecting factual contradictions between new incoming documents and active policies, with visual diffs and 1-click supersede in Review.
drift-arbitrationReal-time verification of context relevance, groundedness, and answer faithfulness with fail-closed hallucination prevention.
quality-gateRun the database-gated suites against test Postgres, including source edit credential preservation, route permissions, and migration-backed behavior.
KVB-012The furthest-out piece of the OKF v0.2 model — trustworthy computation, not just trustworthy prose.
Cross-agent operational ledger with automated memory consolidation and background pattern curation.
agent-memoryKnowledge graph ontology traversals and cross-system entity resolution for multi-hop enterprise reasoning.
graph-ragSanctioned, parameterized computations whose results ship with receipts checkable by deterministic attesters — OKF attestation model for numbers you can trust the same way you trust cited text.
attestation