LlmWikis treats reports, dropped files, AI drafts, and generated summaries as inputs, not automatic public truth. Publication is a promotion decision: useful material moves from source to draft to review to public page only when the right gates pass.
Publication flow
- Intake. Receive source files, reports, screenshots, notes, exported conversations, or generated drafts in the appropriate local bucket.
- Inventory. List files, summarize each source, identify overlap, contradictions, unsupported claims, and useful publication candidates.
- Stage. Convert candidates into reviewable Markdown, page copy, code changes, handoff updates, roadmap items, or security/privacy warnings.
- Review. Check source support, authority boundaries, sensitive data, current-versus-planned support, route placement, and link targets.
- Promote. Seed or update public pages, discovery files, and local handoff state only for claims that survived review.
- Archive. Move processed sources out of active intake after disposition is recorded, leaving pointers for recovery.
Review gates
| Gate | Reviewer question | Blocks publication when |
|---|---|---|
| Source gate | Can the public claim be traced to a source or existing reviewed page? | The claim is unsourced, time-sensitive without date, or copied from a report as fact. |
| Authority gate | Does the page preserve UAIX canonical boundaries for UAI-1, AI Memory, Project Handoff, schemas, registry, and validator behavior? | LlmWikis appears to define or certify UAIX support. |
| Sensitivity gate | Could the content expose secrets, customer data, regulated records, private transcripts, privileged material, or operational access details? | Redaction and review are missing. |
| Support-claim gate | Does the wording separate current support from planned, candidate, internal, or future work? | The page claims public MCP, open editing, automatic publishing, certification, SDK, CLI, or sync before evidence exists. |
| Discovery gate | Should this route be visible in sitemap, llms.txt, navigation, and structured metadata? | The content is private, internal, archived, or still a draft. |
| Verification gate | Were targeted checks run for the files, routes, and discovery records changed? | The change affects public output and no matching local check was run or recorded. |
Safe publication states
- Safe to publish: general architecture, operating patterns, reviewed non-sensitive guidance, and clearly non-normative explainers with sources.
- Needs fact-checking: vendor-specific capabilities, legal or regulatory claims, pricing, schedules, current support matrices, and fast-changing protocol behavior.
- Keep internal: raw intake files, private transcripts, credentials, regulated records, unresolved legal/customer data, and speculative roadmap claims without review.
- Archive after disposition: processed reports, duplicate drafts, stale QA notes, and completed improvement packages.
Human approval boundaries
Human review is required before publishing or changing policy, legal, security, privacy, production, architecture, canonical authority, public support, contributor, or sensitive-data guidance. AI can stage and verify; human governance decides whether the page becomes public reference.