Skip to content

LlmWikis knowledge page

Trust Model

Trust labels tell humans and agents how to use a page. They should be plain, limited, and operational. If a label does not change reader or agent behavior, it probably should not exist.

Label Human meaning Agent behavior Execution and citation rule
authoritative Current source of truth for the stated scope. May cite and rely on it, but still respect sensitivity and human-action boundaries. Execution still requires explicit permission; cite the page and source span.
reviewed Checked by a responsible reviewer but not necessarily the top policy source. Use as dependable context inside its scope. Cite with review date; do not outrank authoritative pages.
dogfood-evidence-backed Derived from private or internal dogfood evidence that passed promotion gates. Use the public lesson, not the private trace. Preserve non-claims. Cite the public promotion route; never imply access to private runtime evidence.
working-draft Useful draft not yet approved. Use as context only; do not present as final policy or decision. No execution; cite as draft when relevant.
proposal Candidate future state. Keep separate from current state and avoid implementing without approval. No execution unless a human explicitly approves the proposal for the task.
needs-review Incomplete, ownerless, weakly sourced, or awaiting reviewer attention. Warn the user and seek better sources or human review. No high-impact action; citation must include caveat.
stale Freshness rule elapsed or source likely changed. Use for background only; verify before operational or public claims. No execution; cite with stale warning and replacement search.
contradicted Known unresolved conflict exists under the stated scope. Surface the conflict rather than choosing a side silently. No execution; cite both sides and ask reviewer.
blocked Privacy, source, legal, safety, authority, or promotion gate blocks use. Stop, name the blocker, and avoid summarizing hidden details. No execution or public citation beyond the blocker record.
historical Past context, not current instruction. Use for background, never as current operating guidance. No execution; route to current replacement if available.
deprecated Superseded and should not guide new work. Do not follow; route to replacement page if available. No execution; cite only to explain history.
external-reference Useful outside source not owned by the wiki. Cite as external context and check canonical source before relying on it. Do not treat as local authority without source-owner review.

OKF validity is not authority

Signal What it proves What it does not prove
type exists The file can be treated as an OKF concept document. The claim is reviewed, current, safe, or source-backed.
okf_version in root index The bundle declares an OKF-compatible entry point. The site is certified, endorsed, or automatically importable.
llmwiki_status The LLMWikis trust label agents should apply. Permission to execute, publish, or overwrite without review.
llmwiki_source_status Whether source support is linked, needed, external, or withheld. That private evidence can be exposed publicly.
llmwiki_agent_use The permitted agent posture for the page. Authorization to cross privacy, production, legal, or destructive-change boundaries.

Retrieval Rank

Rank authoritative and reviewed pages first. Use dogfood-evidence-backed pages when the task is about the promoted public lesson, not when the task requires private runtime facts. Down-rank working-draft, proposal, needs-review, and stale. Stop or ask for review on contradicted and blocked records.

Required Trust Signals

  • Owner
  • Review status
  • Last reviewed date
  • Update frequency
  • Source-of-truth status
  • Confidence or caveat label
  • MATM memory kind, scope, curator decision, and conflict status when the record is shared agent memory
  • Audience and sensitivity
  • Agent permissions
  • Human approval requirement
  • Related authoritative sources