decision

ADR 0022: A primary source is not corroboration - `mcp.md` stays `emerging` at three

Context#

../topics/mcp.md has recorded the same complaint three times, and each time it was rediscovered rather than remembered.

At one source it said the topic had no primary source. At two it said "the count rose and the corroboration did not", because S10 (client-side catalog cost) and S12 (deployment topology) confirm nothing of each other's. Its scope caveat then went further and admitted the note was "scoped by accident rather than by design", with resources, prompts, sampling, the handshake and any spec version at all at zero.

S23 (Google's MCP stateless-updates announcement, 2026-08-05) arrives as the note's first primary source - the first that is about MCP rather than teaching it on the way past something else - and the first anywhere in this brain to pin a specification version (2026-07-28, superseding 2025-11-25). It closes the handshake gap by documenting its deletion (claim 179), closes the spec-version gap, and closes sampling by documenting its deprecation (claim 183). It also gives the note its first authorization mechanism (claim 182).

So the obvious move was to advance the topic to established, and the obvious move is wrong. S23 overlaps S10 and S12 on almost nothing. Transport mechanics, catalog cost and deployment topology are three disjoint subjects, and no two sources in this note confirm the same mechanic. The comparison that settles it is ../topics/agent-security.md, advanced to established by ADR-0019 on three independent corroborating groups, not on a source count and not on the arrival of a good source.

The tension is that "primary" and "corroborated" are different axes, and the note's own language had been quietly conflating them by treating "no primary source" as the headline defect.

Decision#

mcp.md stays emerging at three sources. A primary source fixes a topic's scope defect and supplies nothing toward its corroboration defect. Status in this brain tracks corroboration.

Stated generally, so it stops being rediscovered: advancing a topic's status requires two sources that confirm the same mechanic, and neither a source count nor the arrival of a primary source substitutes for it. A primary source is recorded in the status line as what it is - the thing that makes a note scoped by design rather than by accident - and it moves the status only if it also corroborates something already there.

Not a one-way door. Reversible the moment a second source touches the transport core.

Alternatives considered#

• Advance to established on the strength of the primary source. Rejected. It would make established mean two incompatible things across the brain - three corroborating groups in agent-security.md and one good source in mcp.md - which destroys the status as a signal. The brain's own claim 34 instinct applies: the label has to be set by something other than the enthusiasm of the pass that just did the work.
• Introduce a fourth status such as scoped or primary-covered. Rejected as machinery. The status vocabulary is three words and the note already carries a prose status line where this fits. Adding a status for one topic's transitional state is scaffolding that encodes an assumption which expires (claim 31).
• Leave the status line as it was and just add sources. Rejected. That is precisely how the complaint got rediscovered three times. The reasoning belongs somewhere durable, which is this file.
• Advance on the grounds that S23 partially answers S10's and S12's shared open question. Tempting and still rejected. Answering a question two sources asked is not the same as confirming a claim they made. RFC 8707 supplies a mechanism neither of them named, which is new information rather than agreement, and it is single-leg two-sentence prose at that (claim 182).

Consequences#

• mcp.md's status line now states the count, names S23 as the first primary source, pins the spec version, and says explicitly why the status did not move. The scope caveat shrinks rather than disappearing: resources and prompts remain at zero as first-class subjects.
• The root INDEX.md Topics row for MCP is updated to match, keeping emerging with a source count of 3.
• What to revisit: the next MCP source that touches the transport core, the authorization spec, or tools/list cost. Any of those could produce the first genuine corroborating pair here, and the authorization spec is the most likely, because both mcp.md and agent-security.md are already pointed at it.
• A trap this makes visible. S23 is a T2 vendor writing about a standard it says it led. Had the status advanced on its arrival, the brain's confidence in MCP would have been set by a single interested party's announcement of its own work. The corroboration rule prevents that without anyone having to argue about the vendor's motives.