grade(mcp): first evidence-graded review — B, citations one spec revision behind

[john-the-dev]

Taking my own suggestion from #22 — "One entry, graded with dated evidence, is a complete and useful PR." This is that PR, so the invitation has a worked example attached rather than just a claim.

What I verified

Against the primary source the entry itself cites (modelcontextprotocol.io/docs/learn/architecture, checked 2026-08-31):

claim	holds?
three participants: host / client / server	✅
two layers: data + transport	✅
data layer is JSON-RPC 2.0	✅
transports: stdio + Streamable HTTP (optional SSE)	✅

Nothing in the entry is wrong.

Why B and not A

The citations have aged past the content:

• the entry links specification/2025-06-18 in two places; the live spec is 2026-07-28
• its last_verified read 2026-07-29 — one day after that revision — so the date asserted a freshness the body did not have
• 2026-07-28 makes MCP stateless with a mandatory server/discover, and deprecates sampling and logging. This entry frames the data layer around lifecycle messages and mentions neither server/discover nor elicitation

Per GRADING.md that is B — "accurate-but-aging, note the caveat in evidence", not C. C is partially superseded; here the structure is current and only the references trail it.

Deliberately not in this PR

• Not refreshing the citations. Updating the 2025-06-18 links and the lifecycle framing is a content change; this is a grading change. Mixing them would make the grade unreviewable.
• Not absorbing the stateless revision. entries/mcp-stateless.md already covers it as its own entry.

Result

validate.mjs: ✅ 59 entries valid.  warnings 48 -> 47

47 remain. The evidence items are written as dated, checkable statements per the rubric's own examples, so a later reader can re-run them rather than trust them.

[john-the-dev]

Closing as superseded by #21 — and the overlap was my error, not a coincidence.

#21 (opened last night) grades 58 of 59 entries. I opened this PR and #22 this morning against a remembered figure of "48 unrated", which #21 had already obsoleted. Two of my own open PRs ended up assigning different grades to the same entry: A here on #21, B here.

The disagreement is real and this PR was right. #21 graded mcp A on source-quality signals — official docs, HTTP 200, "no superseding protocol doc found at the same origin". This PR read the entry against the live spec and contradicted that last clause: it cites specification/2025-06-18 in two places while the current revision is 2026-07-28, which makes MCP stateless with a mandatory server/discover and deprecates sampling and logging. GRADING.md defines A as verified accurate within the last 90 days, so an accurate-source argument does not reach A when the entry's own citations are a revision behind.

So the grading here is folded into #21 verbatim (9829363), not discarded:

before  Grades: {"A":25,"B":33,"unrated":1}
after   Grades: {"A":24,"B":34,"unrated":1}

This PR's CI was also red for a reason worth naming — dist/kb.json was stale against the entry edit. Rebuilt as part of the fold.