This is the practical LlmWikis profile for a source-governed, AI-navigable wiki. It defines the minimum contracts that make a wiki safe for humans, search, retrieval, handoff, and agent-assisted maintenance. It is not a universal standards claim and it does not replace UAIX for UAI-1, AI Memory, Project Handoff, schemas, registries, validators, or conformance language.
Specification boundary
Use these requirements when a project wants to claim it follows the LlmWikis handbook profile. Use UAIX.org for canonical UAI-1 and Project Handoff definitions. A project can adopt part of this profile without claiming certification, endorsement, hosted validation, or automatic write support.
Profile contracts
| Contract | Required behavior | Failure if missing |
|---|---|---|
| Layer contract | Separate immutable raw sources, reviewed wiki pages, index/log navigation, and optional derived graph/search/cache artifacts. | Agents cannot tell evidence from compiled truth. |
| Page contract | Every durable page has title, canonical URL, owner, status, source status, last reviewed date, sensitivity, agent-use rule, and source trace when it makes factual claims. | Search and agents over-trust fluent but unowned pages. |
| Navigation contract | Index, log, breadcrumbs, children, typed related links, aliases, redirects, sitemap, and llms.txt agree on canonical routes. | Readers land on duplicates, stale pages, or orphaned evidence. |
| Trust contract | Authoritative, reviewed, draft, stale, contradicted, deprecated, archived, and blocked states change reader and agent behavior. | Status labels become decorative and unsafe. |
| Source contract | Raw evidence remains preserved or linked; compiled claims cite source pages, source records, or canonical external routes. | Useful synthesis becomes unverifiable. |
| Agent contract | Agents read the smallest useful page set, cite pages, stage risky updates, preserve uncertainty, and stop on authority, privacy, security, legal, production, or destructive-change boundaries. | Agent work becomes silent rewriting. |
| Search contract | Search results expose canonical pages, descriptions, freshness/status signals where available, and route users back to reviewed pages before retrieval or graph layers. | Search becomes a prettier document dump. |
| Validation contract | Lint blocks duplicate canonicals, broken links, private discovery leaks, missing required metadata, unresolved unsafe claims, and generated artifacts posing as authority. | Bad structure compounds across sessions. |
Conformance profiles
| Profile | Minimum artifacts | Appropriate claim |
|---|---|---|
| Profile A: file-native wiki | raw/, wiki/, wiki/index.md, wiki/log.md, README, source policy, and owner/status metadata. |
The project has an exportable source-backed wiki skeleton. |
| Profile B: public handbook wiki | Profile A plus canonical public routes, sitemap, llms.txt, metadata descriptions, redirects, noindex/archive policy, and private-file exclusion. | The public surface is crawler-readable and source-aware. |
| Profile C: agent-ready wiki | Profile B plus AGENTS.md or CLAUDE.md, agent read order, update rules, citation rules, stop conditions, staged-write policy, and review checklist. | Agents can use the wiki under explicit boundaries. |
| Profile D: graph/retrieval-ready wiki | Profile C plus stable IDs, typed relations, claim/source-span policy, contradiction records, graph lint, and retrieval/abstention rules. | Search, graph views, or retrieval caches can be derived from reviewed pages. |
| Profile E: handoff-ready wiki | Profile D plus hot/cold memory boundary, AI Memory or Project Handoff pointers when used, transfer evidence log, and archive manifest discipline. | Durable wiki memory can safely feed compact handoff context. |
Required artifact set
| Artifact | Role | Validation question |
|---|---|---|
README.md |
Front door for scope, owner, audience, and first route. | Can a new reader tell what the wiki covers in one minute? |
raw/ |
Immutable source evidence, exports, summaries, checksums, and manifests. | Can a claim be checked without relying on chat memory? |
wiki/ |
Reviewed concept, guide, reference, runbook, decision, source, and contradiction pages. | Are durable answers stored as inspectable pages? |
wiki/index.md |
Route map for hubs, page types, owners, statuses, and high-value tasks. | Can agents route before broad reading? |
wiki/log.md |
Append-only ingest, query, lint, review, correction, and archive record. | Can future work see what changed and why? |
AGENTS.md or CLAUDE.md |
Agent operating contract, read order, allowed paths, stop conditions, and checks. | Can an agent avoid guessing permissions? |
TRUST_MODEL.md |
Status labels and their human/agent meaning. | Does each label change behavior? |
SOURCE_POLICY.md |
Source admission, citation, redaction, review, and canonical authority rules. | Can weak or private sources be stopped? |
page-registry |
Canonical URL, aliases, redirects, page type, owner, status, and discovery state. | Can duplicate canonicals be detected? |
llms.txt and sitemap.xml |
Public AI/search discovery for canonical routes only. | Are private files, archives, and generated package mirrors excluded? |
Minimal page record
---
title: "Deployment Process"
canonical_url: "https://example.com/operations/deployment-process/"
description: "Reviewed deployment workflow, approvals, rollback path, and agent boundaries."
page_type: "runbook"
owner: "Platform Team"
status: "current"
source_status: "source-linked"
last_reviewed: "2026-05-24"
review_cycle: "after each release process change"
sensitivity: "internal"
agent_use: "read-cite-propose"
source_trace:
- "raw/releases/2026-05-process-review.md"
typed_relations:
part_of:
- "/operations/"
depends_on:
- "/reference/security-privacy/"
supersedes: []
contradicts: []
source_for: []
---Search and retrieval requirements
- Search should prefer canonical reviewed pages over aliases, archive pages, raw files, package mirrors, and generated history.
- Result cards should expose a useful title, short description or excerpt, route, and freshness/status signals when the platform can show them.
- Search should help users recover from broad queries with route shortcuts for specification, metadata, repair, source policy, AI Memory, and the setup wizard.
- Retrieval, vector search, graph traversal, or GraphRAG should consume reviewed wiki pages and source summaries, not private intake folders or raw unreviewed dumps.
- Answers produced from search should cite the pages or source records used, preserve contradictions, and abstain when evidence is missing, stale, private, or blocked.
Validation gates
| Gate | Blocks publication when | Typical check |
|---|---|---|
| Discovery gate | Private files, archives, admin URLs, package mirrors, duplicate aliases, or noindex drafts appear in public discovery. | Compare sitemap, llms.txt, robots, public routes, and route registry. |
| Metadata gate | Required page fields are missing or invalid. | Validate frontmatter/page registry against the metadata standard. |
| Canonical gate | Two pages claim the same concept without a winner, redirect, or archive decision. | Duplicate title, duplicate intent, and duplicate canonical URL scan. |
| Source gate | Factual claims lack source trace, canonical external route, or review status. | Claim/source-span and source-policy lint. |
| Safety gate | Secrets, private data, regulated material, or privileged operational detail can leak to public pages or portable packets. | Secret patterns, sensitivity labels, redaction review, and export review. |
| Agent gate | Permissions, stop conditions, destructive-change rules, or targeted checks are unclear. | AGENTS.md/CLAUDE.md plus task-specific verification checklist. |