Skip to content
Aftergraph Knowledge Plane

Trust & Verification

Trust here is mechanical, not rhetorical. Every page carries the receipts.

What the plane guarantees (structurally)

Section titled “What the plane guarantees (structurally)”
  1. Ownership is exact. Every page binds to owner-repo@<full-sha>. Rendered views cannot outlive their provenance.
  2. Private content cannot leak. A build-time allowlist + a scan of every served byte fail the build on any private-repo reference.
  3. Claims are verbatim. The claim register is copied from the ISR evidence audit — statuses, limitations, reproduction states. The plane renders claims; it never strengthens them. Projection can only weaken (live REVERSED → CONTESTED), never upgrade.
  4. Staleness is visible. Freshness verdicts distinguish harmless repository churn from real semantic drift; UNKNOWN fails honest, never optimistic (ADR-004).
  5. Agents get the same truth. The Docs MCP is read-only, build-bounded, and carries site_commit in every answer.

The nine gates

Section titled “The nine gates”

npm run build runs them all; CI runs them remotely; the smoke suite runs against production after deploy:

# Gate Catches
1 Source manifest schema Non-full SHAs, non-public visibility
2 Provenance coverage Pages without owner bindings
3 Internal links Dangling references
4 OpenAPI validity Broken/absent canonical API artifact
5 Ownership allowlist Cross-registry contamination
6 llms.txt presence Missing agent surface
7 Artifact fingerprints Tampered/uncertain consumed artifacts
8 Graph consistency Edges to unknown nodes, non-canonical claim states
9 Context packs Packs missing provenance/freshness bounds

What the plane will not do

Section titled “What the plane will not do”
  • Upgrade, soften or reinterpret a claim — the registry is authoritative.
  • Auto-pin a moved source — adoption is a human-reviewed commit.
  • Serve a fresher view than its build — every surface states its boundary.
  • Host canonical specs or claims — owning repositories do that.

Verify us, don’t trust us

Section titled “Verify us, don’t trust us”
  • Reproduce the freshness verdict: node scripts/freshness.mjs against live GitHub state.
  • Diff a claim against the registry at the pinned commit (Tutorial 1).
  • Point an MCP client at the plane and query it yourself (Tutorial 2).

Contradictory prior art or a failing invariant? Open an issue on Aftergraph/docs — criticism that can falsify a claim is worth more than agreement that cannot.

Was this page helpful?