decision

ADR 0014: No new topic for organisational context - it is context engineering with a different corpus

Context#

S11 (How we built LangChain's agent-first data stack, 2026-07-27) is the first source in this brain whose subject is an agent reading a company's own proprietary knowledge - metric definitions, business rules, which dashboard is canonical - rather than reading a codebase, a tool catalog, or its own memory.

That is a genuinely new corpus, and it came with a vocabulary the existing notes do not use: semantic layers, data contracts, endorsements, dbt models, workspace guides. The deep-research pass (R2) added more: knowledge engineering, the knowledge acquisition bottleneck, authority control. A recognizable, reusable area with an established name in an older discipline is exactly what AGENTS.md §"Scope: topics are open" says to capture as a new topic.

So the question was real: does organisational-context (or knowledge-engineering, or data-context) deserve its own note?

Against it stands the counter-rule in the same section - "don't spawn a topic per source" - and the fact that S11's claims, once stripped of their domain vocabulary, kept landing in notes that already existed.

Decision#

No new topic. S11's claims are distributed across three existing notes:

The deciding argument: nothing in S11 is specific to data. Re-read claim 92 with "table" replaced by "API endpoint", or claim 95 with "dashboard" replaced by "internal wiki page", and both hold unchanged. The five-store decomposition is a way to organise any proprietary context an agent must read. A new note would have collected the domain vocabulary while the transferable claims stayed elsewhere - which is the worst outcome, because the vocabulary is the disposable half.

This is not a one-way door. Opening the note later costs three cut-and-pastes and an INDEX row.

Alternatives considered#

• Create organisational-context.md. Rejected on the "one topic per source" guardrail, and because every claim already had a better home. It would have started life with one source, one vendor, no measurements, and heavy overlap with two established notes - and rag.md already carries a standing watch item about overlapping with context-engineering.md. Adding a third note that straddles both would make that worse.
• Create knowledge-engineering.md on the strength of the R2 F5 hop. Rejected as premature and backwards: the hop is real and valuable, but it arrived from a research pass on one source, not from a source about knowledge engineering. Promoting a research finding into a topic would invert the evidence order this kit runs on. It is recorded as claim 98 in rag.md, where it strengthens claim 72.
• File everything under rag.md. Rejected because S11's context layer is not retrieved in the RAG sense - four of its five stores are resident context assembled per query by the agent's host, and the note would have had to redefine its own scope to absorb them.
• File everything under agents.md. Rejected: S11 says nothing about the loop, control flow, tools or orchestration. It was listed as a topic on the source folder during ingest and removed before compounding for exactly this reason.

Consequences#

• Easier: claim 93 now reads as a three-source pattern in one place, which is the strongest claim S11 produced and would have been split across two notes otherwise. Claim 99 (the head/tail inversion) only works because S10 and S11 sit in the same note.
• Harder: context-engineering.md grows again - it is now 6 sources and the longest note in the brain. This is the note to watch for a split, and the natural seam if one is ever needed is authoring context (prompts, definitions, contract documents) versus managing it at runtime (compaction, middleware, budgets). Not yet: the seam is not load-bearing while the same claims inform both.
• skills.md gained an S11 row but no claims, and its stale "1 source" status line was corrected to 4 in the same pass. Note the trap that correction guards against: the count rose from 1 to 4 while the evidence did not move at all, because S7, S9 and S11 each contribute one peripheral observation and none studies skills. The status line now says so explicitly.

Revisit when: a second source arrives whose subject is an agent over a company's proprietary non-code knowledge - a data-catalog, enterprise-search or internal-documentation source. At two sources the vocabulary starts to corroborate rather than merely accumulate, and the split argument becomes real. Until then, the claims are where a reader will look for them.