Skip to content
Aftergraph Knowledge Plane

Tutorials

Learning-oriented. Each tutorial is a complete, runnable walkthrough — start with Quickstart if you have not done the 10-minute loop yet.

Tutorial 1 — Trace a claim to its registry commit

Section titled “Tutorial 1 — Trace a claim to its registry commit”

Goal: take a claim rendered on this portal and verify it, character for character, against the authoritative ISR registry at the pinned commit.

  1. Open the Claim Graph and pick a claim — say C-002 (“Control plane token footprint is capped under 300 tokens”). Note its Source SHA (72e115c6…).

  2. Fetch the authoritative registry at exactly that commit:

    Terminal window
    gh api "repos/Aftergraph/intelligence-systems-research/contents/EVIDENCE-AUDIT-AND-CLAIM-REGISTRY.md?ref=72e115c642441e10dd88890edfac86b787315cf0" \
      --jq '.content' | base64 -d | grep "C-002"

  3. Compare: the portal’s auditStatus for C-002 is VERIFIED — the same string the registry row carries. The plane renders, never reinterprets. If they ever diverge, the registry wins and the plane’s build is stale.

  4. Check the freshness of that pin:

    Terminal window
    curl -s https://docs.aftergraph.org/status.json | python -m json.tool | grep -A4 intelligence-systems-research

What you learned: every claim on this portal resolves to a byte-exact registry row at a pinned commit, and the freshness system tells you whether that pin still matches live remote state.

Tutorial 2 — Consume the plane via MCP

Section titled “Tutorial 2 — Consume the plane via MCP”

The plane ships a read-only MCP server (scripts/docs-mcp-server.mjs) that serves the built site. Wire it into any MCP client:

  1. Build the site locally (the MCP serves the built artifacts):

    Terminal window
    npm run build

  2. Register the server (Hermes example):

    mcp_servers:
      aftergraph-docs:
        command: node
        args: ["/absolute/path/to/docs/scripts/docs-mcp-server.mjs"]

    Claude Code and other stdio-MCP clients use the equivalent command/args shape in their own config.

  3. Verify the tool surface:

    Terminal window
    printf '{"jsonrpc":"2.0","id":1,"method":"tools/list"}\n' | node scripts/docs-mcp-server.mjs

    Seven tools: docs_get_status, docs_get_contract, docs_get_claim, docs_get_page, docs_search, docs_get_catalog, docs_get_context_pack.

  4. Ask it something real:

    Terminal window
    printf '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"docs_get_claim","arguments":{"id":"C-008"}}}\n' \
      | node scripts/docs-mcp-server.mjs

What you learned: agents consume the same provenance-bounded truth as humans — every MCP answer carries the build’s site_commit, so an agent can never mistake a rendered view for a fresher source.

Was this page helpful?