Governance keeps the public handbook useful without overstating support. These rules define what can be published now, what needs review, what must stay internal, and what cannot be claimed until the supporting system actually exists.
Editorial rules
Use source-backed writing, reviewer status, clear update dates, visible uncertainty labels, and non-normative labels on UAIX-related explainers.
Operator-owned decisions
License choice, privacy policy, terms, DMCA/takedown, contributor agreement, analytics disclosure, and moderation escalation require human approval before public intake expands.
Publication is gated
Raw sources, dropped improvement files, AI drafts, and staged Markdown are not public truth by default. Promote them only after source review, authority checks, sensitive-data review, and targeted verification. See Review-Gated Publication Model.
Current operating model
| Area | Current state | Boundary |
|---|---|---|
| Public handbook | Live pages can teach LLM Wiki architecture, operations, schemas, templates, examples, and source policy. | Do not publish private intake files or internal audit pages as public content. |
| UAIX and UAI-1 | LlmWikis can explain and compare as a non-normative case-study layer. | UAIX.org remains canonical for specs, schemas, registry, validator behavior, roadmap, governance, and Project Handoff. |
| Contributions | Controlled source suggestions, bug notes, and reviewed batches are acceptable. | No anonymous public editing until legal, privacy, moderation, and abuse handling exist. |
| AI assistance | Useful for drafting, summarizing, structuring, and linting. | AI output is not reviewed public truth until source and boundary checks pass. |
| Automation | Local helpers and future private tooling may support review. | No live benchmark, automated ingestion, public MCP, certification, or multilingual claim until implemented and reviewed. |
Maintenance model for an LLM Wiki
| Practice | Minimum rule | Why it matters |
|---|---|---|
| Page ownership | Every authoritative page has a named owner role and review cycle. | Agents and readers can route questions to accountable humans. |
| Human review of AI edits | AI may draft or stage updates; policy, security, privacy, legal, architecture, and production pages require human approval. | Prevents fluent AI edits from silently becoming authority. |
| Staleness detection | Pages past review cycle move to needs-review until checked. | Old confident pages stop misleading readers and retrieval systems. |
| Deprecation and archive | Deprecated pages point to replacements; historical pages say they are not current guidance. | Old context remains useful without becoming current instructions. |
| Incident-driven updates | Incidents update runbooks, decisions, open questions, and affected system pages. | The wiki learns from operations instead of preserving stale procedures. |
| Link and metadata audits | Run recurring checks for broken links, missing owners, stale dates, duplicate pages, and missing sensitivity labels. | Structure decays unless maintenance is routine. |
Risk register
| Risk | Mitigation | Escalate when |
|---|---|---|
| Normative confusion | Place non-normative labels on UAIX explainers and link to canonical UAIX routes. | A page implies LlmWikis defines UAI-1 behavior. |
| AI content drift | Require human review, source anchoring, and status labels before public reference status. | A draft makes factual claims without sources or dates. |
| Low-quality editing | Start invite-first; add templates, review queues, and contributor roles before open editing. | A submission changes many topics or mixes unrelated fixes. |
| Legal exposure | Choose license, privacy, and takedown workflows before accepting public submissions. | A contribution includes copyrighted, personal, confidential, or restricted material. |
| Public/internal mix-up | Keep Content, Improvement, and Archive buckets distinct and keep audit reports out of public navigation. | A chore list, SEO report, QA note, or internal backlog route appears publicly. |
Publishing gates
- Source gate: time-sensitive model, benchmark, law, provider, and protocol claims need source links and dates.
- Authority gate: UAI-1 pages must link to UAIX for canonical specifications, schemas, registry, validator behavior, roadmap, and governance.
- AI gate: AI-assisted text can help draft, but human review decides whether the page becomes public reference.
- Contribution gate: public editing stays closed until contributor, abuse-handling, privacy, and legal workflows are approved.
- Discovery gate: public sitemap, llms.txt, header, footer, and structured data must not list private handoff files, archived intake, generated packages, or retired audit routes.
- Retirement gate: pages removed for being internal work should return a permanent gone response or be clearly drafted before release.
Review-Gated Publication Model
How evidence and staged drafts become public pages without turning intake files into automatic truth.
Source Policy
Status labels, citation rules, AI-assisted draft boundaries, and stale-content handling.
For AI Agents
Agent reading order, citation expectations, uncertainty handling, and approval boundaries.