One record,
every surface in sync.
Aplinode’s knowledge lives inknowledge/as plain markdown. The same file is the website, the agent surface, the search index, and the audit trail. Edit the record, and every surface updates on the next build.
0
0
0
4
Live record
Aplinode
Docs
0
Stable
0
Audit
On
What this is
A knowledge system, not a website.
Most websites are a copy of a document. This one is the document. Every page, every agent answer, and every search result is derived from the same governed markdown inknowledge/.
The record
Plain CommonMark files. No MDX, no framework, no JSX in the source. The file is the artifact.
Governed
Status, audience, owner, approval, takedown — every fact is frontmatter the site renders verbatim.
Agent-ready
An MCP door, an llms.txt, and per-document twins — the same record, surfaced for machines.
Versioned
Each ingest mints a generation. The previous one stays as a rollback target until you collect it.
Surfaces
One record, four faces.
Each surface reads the sameknowledge/. Build decisions are made once, by staging, and read back by every surface that serves the result.
01 — Human surface
The website
A static, statically searchable site with a sidebar, breadcrumbs, prose styling, and per-document audit. Every page carries its owner, effective date, and read time on the page.
02 — Agent surface
MCP + llms.txt
An MCP door returns cited, honest retrieval with a calibrated abstention floor. llms.txt and per-document .md twins are the same record, ready for any LLM to ingest.
03 — Search index
Client-side search
A prebuilt Orama index ships with the build, so search runs in the browser over a static index — no server, no rate limit, no third party.
04 — Build artefact
Static output
A reproducible, byte-stable build. Commit build.lock.json: it is the provenance every artefact stamps, and the takedown ledger is checked against it on every refresh.
Governance
Truth has a trust tier.
Every document says where it came from, who approved it, and who may read it. The site renders those facts — it does not invent them. Nothing reaches a surface it has not been admitted to.
Trust tiers
Verified claims, machine-checked provenance.
- human-reviewedA human actor has read and approved this document.
- machine-confirmedCross-checked by an automated process; no human has approved it.
- unverifiedA draft, or a claim nobody has reviewed. The honest word for it.
Tiers live in verified on each document. The agent surface admits nothing below the calibrated floor.
Lifecycle
draft → stable → deprecated.
- draftAdmitted to no machine surface. Visible in the dev preview only.
- stableAdmitted everywhere. Carries a generated stamp and an approval actor.
- deprecatedReplaced by a successor. Withdrawn from the agent surface; visible on the site with a supersession notice.
Takedowns are an append-only ledger (.ksor/takedowns.yaml), not a deletion — the record is what an agent can cite.
The record
What’s inside.
The map of every document this record holds, in the reading order the sidebar and llms.txt agree on.
npm run check after authoring the first document underknowledge/.Two readers, same record
For agents and for humans.
For agents
Cited retrieval, honest abstention.
POST /mcp returns answers with provenance and a calibrated abstention floor. When a question is outside this record, the door says so — rather than confabulating from adjacent context.
- Per-document stable IDs (
knowledge/<path>) so citations survive a rebuild. - Discovery document at
/.well-known/oauth-protected-resource/mcpwhen SSO is configured. - JWT audience-enforced. A token minted for another service is rejected, not replayed.
For humans
The record, on the page.
Every page carries its own provenance: owner, effective date, read time, and a chip if it is a draft or has been replaced. Sources are cited inline, not bolted on at the end.
- Three voices — display for the record, sans for the site, mono for the machine.
- Press Ctrl + K to search across every document.
- Summaries, slides, flashcards, and quizzes ride along on the documents they belong to.
The record is built from governed markdown, served as a static site, surfaced as a search index, and exposed to agents over an MCP door. Every surface is a derivation. Edit the record, and the next build updates everything else in lockstep.