Skip to content
Aftergraph Knowledge Plane

Explanation

Why a compiler, not an owner

Section titled “Why a compiler, not an owner”

Documentation drifts because portals copy truth. When a spec moves, the copy stays — silently wrong. The Knowledge Plane inverts the model: repositories own truth, governance owns boundaries, evidence owns claim strength. This plane only resolves pins and renders. It cannot be silently wrong about ownership because it ships the exact SHA it rendered from — and its build fails if that SHA leaves the public allowlist or a link dangles.

Full boundary decision: ADR-002 — source-of-truth boundaries.

The four clocks

Section titled “The four clocks”

Any statement about a system answers four different time questions, and conflating them is how documentation lies:

  1. Source truth — which commit contains this artifact? (pinned SHA)
  2. Semantic truth — which spec/API version does it describe? (e.g. OpenAPI 3.1.0 / API v0.2.0, Draft 0.3)
  3. Evidence truth — which evidence cut supports the claim? (e.g. AUDIT-EVID-001 2026-09-04)
  4. Live truth — what is the remote state right now? (freshness verdict)

Every page, status surface and context pack states which clock it speaks from. A page can be CURRENT in source terms while its claim’s evidence cut is months old — that is not a contradiction, it is the design.

Repository movement is not semantic movement

Section titled “Repository movement is not semantic movement”

The first live verification taught this: several pinned repos had advanced (dependabot, feature commits) while the Work Intelligence OpenAPI artifact — the thing this plane actually consumes — was byte-identical. Treating git movement as documentation drift produces alarm fatigue; treating pins as eternal produces silent rot. The policy splits the two: SOURCE_MOVED_CONTENT_UNCHANGED for harmless churn, SOURCE_MOVED_CONTENT_CHANGED escalating to STALE for real drift, and adoption is always an explicit human-reviewed commit (ADR-004).

Why claims carry two statuses

Section titled “Why claims carry two statuses”

The ISR registry speaks in audit language (“STRONGLY_SUPPORTED (SAMPLE-BOUNDED)”, “REVERSED”, “UNTESTED”). Agent-facing tooling needs a closed vocabulary (SUPPORTED / PARTIALLY_SUPPORTED / CONTESTED / REFUTED / OBSOLETE). The plane projects one onto the other and keeps both: the canonical state for filtering, the verbatim audit wording for truth. The projection can only weaken confidence (live REVERSED → CONTESTED), never strengthen it. See ADR-005.

Why read-only MCP

Section titled “Why read-only MCP”

Docs surfaces that let agents write create a second, unofficial source of truth. The Docs MCP is structurally incapable: it reads built artifacts, has no shell, no network, no write path — and every answer carries the site_commit it was served from, so consumers can never over-trust freshness.

Was this page helpful?