decision

ADR 0013: Secondary but substantial - a source advances a topic it is not about

Context#

ADR-0012 ruled that a mention is not a source, and closed by naming the case it had not yet produced:

Revisit when: ... a source is genuinely borderline - substantial on a topic but entirely secondary to its main subject - which this pass did not produce.

S10 (Tool search, Microsoft, 2026-07-29) produced it, one source later.

S10 is about tool retrieval. Its thesis is that a large tool catalog is an information-retrieval problem; its measurements are token counts and Recall@10. It is not an MCP article. MCP is never its subject, and a reader looking for a protocol explainer would not choose it.

And yet it teaches more about MCP mechanics than anything else in this brain, because its mechanism is built directly on the protocol's shape:

• tools/list is a per-turn cost centre whose size scales with the catalog, not the task (n1, corroborated by Figure 1 and Figure 2).
• Runtimes reject calls to tools absent from tools/list, which is the entire reason the design needs a second proxy tool rather than one (n4). That is a statement about how MCP clients behave, not about Foundry.
• An MCP server can front other MCP servers, exposed over streamable HTTP with bearer auth, with the aggregated servers appearing as ordinary attached MCP endpoints (n7, from the portal screenshot).
• A significant capability was added with no new protocol primitive - stated as a design goal and visible in the product surface as a toggle (n5).
• What a tool exposes to an index - name, description, argument names and argument descriptions, three levels deep - and the fact that a field can be indexed while staying invisible in MCP responses (n8, n14).

Under ADR-0012's own test - would a reader arriving at this note from the INDEX find what the status promises? - a mcp.md carrying those five items answers real questions about how agents consume MCP tools at scale. That is not a labelled box in a diagram.

mcp.md has been seed with zero sources since this brain started, so the call is also the one with the most inertia behind it, in both directions.

Decision#

A source advances a topic's status when it teaches that topic's scope substantively, regardless of whether the topic is the source's subject. Being secondary is not disqualifying; being thin is.

Applied this pass: mcp advances seed to emerging on S10, with its Synthesis section written from S10's gated nodes and nothing else. S9's earlier sighting stays recorded as a sighting and is not retroactively promoted to a source - it contributed one box in one diagram and that has not changed.

The test is unchanged from ADR-0012; only the qualifier is removed. Scope is defined by the note's own "What this covers" section, and the question is whether the note can now answer a basic question about the area. What was implicit before and is now explicit: that question is about the note, not about the source's table of contents.

Why "secondary" was the wrong axis#

ADR-0012 was written against over-counting, and it worked. But it left "substantial" and "primary" tangled together, and they measure different things:

The bottom-right cell is what ADR-0012 governs and it still governs it. The top-right cell had no rule, and defaulting it to "sighting" would have been the under-counting error - which ADR-0012 described as cheap and self-correcting, and it is, but only until it starts systematically discarding the sources a topic actually has. Secondary sources are where cross-domain learning lives. A brain that only counts a source for the topic on its title page will never notice that the tool-retrieval article is its best MCP source, or that the memory article is also a concurrency source.

The failure this does not reopen. ADR-0012's real target was a status advancing on a word appearing, leaving a populated status row over an empty Synthesis section. The guard against that is unchanged and is now doing all the work: the Synthesis section must be written, from gated nodes, before the status moves. If a source cannot fill it, the source did not teach the topic, whatever its subject was.

Alternatives considered#

• Hold mcp at seed until a protocol-first source arrives. Rejected. It would leave five cited, corroborated claims about MCP mechanics sitting in a source note with no topic to feed, and it states something false: the brain would claim to know nothing about MCP while holding a measured account of what tools/list costs. The honest version is emerging with a scope caveat, which is what the note now carries.
• Advance mcp and also count S9 retroactively, taking it straight to established. Rejected outright - that is precisely the ADR-0012 error, and doing it in the same pass that cites ADR-0012 would be self-refuting. S9 still contributed one box.
• Create a tool-use topic and put everything there instead. Rejected as topic-per-source (AGENTS.md: "Don't spawn a topic per source"). S10's claims already have four good homes, and the IR-flavoured ones belong in rag.md where the retrieval machinery lives. Revisit if a second and third source arrive on tool selection specifically, at which point splitting a tool-use note out of agents.md and context-engineering.md becomes a real question rather than a speculative one.
• Require two secondary sources to equal one primary. Rejected as arithmetic standing in for judgement, on ADR-0004's line: validate.py checks form, the architect judges scope. It would also have no principled exchange rate.

Consequences#

• Easier: a topic can now be advanced by the source that genuinely taught it, which is frequently not the source that advertises it. Expect this to matter most for the thinnest topics, since a well-covered topic gets primary sources anyway.
• Harder: the judgement is now unavoidable on every ingest of a broad source, in both directions - ADR-0012 asks "is this only a mention?" and this asks "is this actually substantial?". The mitigation is the same and it is cheap: try writing the Synthesis section. The attempt answers the question.
• A new drift risk, and it is the mirror of ADR-0012's. A topic whose entire content arrives sideways from sources about other things can end up scoped by accident - covering exactly the slice its secondary sources happened to touch, with the caveat that says so slowly going unread. mcp.md is now in that state by construction and says so in its status line. The dream pass should treat a topic with no primary source as a watch item, alongside ADR-0012's accumulated sightings.
• Revisit when: a topic reaches established without a single source that is about it (does the ladder need to distinguish coverage from breadth?), or when a secondary source's claims turn out to be systematically narrower than the note's stated scope.
• Follow-up edits: mcp.md - status, Synthesis, Key claims, Sources feeding, and the S9 sighting rewritten to point here; INDEX.md Topics row for MCP; claims.md claims 83, 84, 89.