decision

ADR 0012: A mention is not a source - naming a topic does not advance its status

Context#

Ingesting S9 (Inside the Microsoft Agent Framework, 2026-05-28) forced the same call twice in one pass, on two different topics, and the kit had no written rule for it.

The mcp case. S9's ecosystem figure names MCP as one of exactly two tool-integration standards, beside OpenAPI [S9 fig_AgentLoop]. brain/topics/mcp.md has been at seed with zero sources since the brain started. On a literal reading of the status ladder - seed (created, no source) -> emerging (one source) - S9 is a source that mentions MCP, so mcp advances and the brain's longest-standing empty topic is finally populated.

The skills case. S9's harness inventory files Skills under Context, beside Prompts and Memory [S9 fig_AgentHarness]. skills.md is at emerging on S5, with S7 adding the category name (claim 64). A third source touching skills would, read literally, be the second corroborating source that takes the topic to established.

Both readings are wrong, and wrong in the direction that inflates the brain. What S9 supplies in each case is one box in one diagram. It never defines MCP, never describes a server, tool, resource, prompt or transport, and never mentions MCP in its prose at all. It never defines a skill, never discusses writing, triggering, evaluating or retiring one, and never mentions skills in its prose either. A topic note that advanced on this would announce coverage it does not have - and mcp.md would go to emerging while its Synthesis section stayed literally empty.

This is a taxonomy question, which is why it is an ADR and not a judgement buried in a log entry: it will recur on every broad source, and broad sources are common. The status ladder in AGENTS.md counts sources, and until now "source" was undefined at the low end.

Decision#

A source advances a topic's status only when it teaches something about that topic's scope. A mention, a citation, or a labelled box in a diagram is a sighting, and a sighting is recorded in the topic's Open questions without advancing status or appearing in "Sources feeding this topic".

Applied this pass:

• mcp stays seed. The sighting is recorded in that note's Open questions, with an explicit line saying it is deliberately not counted and why. "Sources feeding this topic" stays empty.
• skills stays emerging. S9 is listed under Sources feeding, because its box placement does bear on a live claim (64) - but the entry states plainly that it contributes one box, supports the family assignment only, and does not move the status.

The test: would a reader arriving at this note from the INDEX find what the status promises? If the status says emerging and the note cannot answer a basic question about the area, the status is lying. Scope is defined by the note's own "What this covers" section, which is what makes this checkable rather than a matter of taste.

Not a one-way door. Nothing is discarded - the sighting is written down, cited and findable. If a later source makes the area real, the sighting is already there to cite alongside it.

The asymmetry that decides it#

The two errors are not symmetric, which is the same argument ADR-0006 makes about the static probe.

Over-counting also corrupts the one thing the ladder exists to measure. If established can be reached by three sources each naming a topic in passing, it stops meaning "this brain knows this area" and starts meaning "this string has appeared three times".

Alternatives considered#

• Count any citing source, and let confidence carry the weakness. Rejected: status and confidence measure different things and this conflates them, in the opposite direction from ADR-0008. There, status advanced while confidence deliberately did not, because the area genuinely recurred across two sources. Here the area does not recur at all; only the word does. Confidence columns cannot rescue a status that is false.
• A numeric threshold (N claims promoted, or a word count). Rejected as the wrong instrument, on ADR-0004's line: this is a judgement, and encoding it as a count invites gaming and breaks on the first source that teaches a great deal in two sentences. validate.py checks form; the architect judges scope.
• Advance mcp to emerging and write the Synthesis section from general knowledge. Rejected outright, and worth recording because it is the tempting one. It would produce an uncited topic note - a direct violation of ground-every-claim - and it is precisely the failure S8's ingest already caught once: a claim written about a source that had not been ingested, wrong the day it was written, in a file nothing re-reads. The brain's value is that everything in it came from somewhere.
• Drop the sighting entirely. Rejected: the fact that a major vendor treats MCP as one of two tool-integration standards is real information about MCP's position in the ecosystem. It is simply not information about MCP's mechanics. Recording it in Open questions keeps it findable without letting it masquerade as coverage.

Consequences#

• Easier: topic status becomes trustworthy again. A reader can take emerging to mean the note has content, and the INDEX stops being able to advertise an empty note.
• Easier: broad, ecosystem-shaped sources (framework announcements, survey posts, landscape talks) can now be ingested without dragging every topic they name up the ladder. Expect more of these, not fewer.
• Harder: every ingest of a broad source now carries a per-topic judgement call, and the architect has to make it explicitly rather than by counting. The mitigation is that the call is cheap and the test is written down.
• A new place to look for drift. Sightings accumulate in Open questions sections and nothing sweeps them. The dream pass should treat an accumulation of sightings on one topic as a promotion signal - three sightings and no source is itself evidence the area is real and under-covered. Added as a watch item rather than a rule, since it has not happened yet.
• Revisit when: a topic accumulates sightings but never a real source (does the ladder need a rung between seed and emerging?); or a source is genuinely borderline - substantial on a topic but entirely secondary to its main subject - which this pass did not produce.
• Follow-up edits: mcp.md Open questions + Sources feeding; skills.md Sources feeding; S9's LEARNING.md "Feeds these topics"; Topics row in that source's SOURCE.md (which named evals for a topic nothing was promoted to - the same defect class, caught on review); no INDEX Topics-row changes, which is the point.