ADR-183: The AI section exists, and editor surfaces live in it 

Context 

ADR-119 decided Administration, for now, and left the placement open with a named revisit trigger: the first cross-consumer editor surface. Issue #812 carried that open question so the next surface would not settle it by accident.

The question was put and answered rather than triggered. The section is created, and it is called AI.

Two arguments were weighed once more before deciding, and both are recorded because the second overturns a reason ADR-119 gave for the name.

"LLM" is the more precise name. It is not. Of the four products that share the section, nr_repurpose generates audio and images, nr_ai_search fuses BM25 keyword search with embeddings, and the cowriter drives CKEditor — only nr_llm is a language-model product. A section named LLM holding a product called AI Search that runs no language model would be wrong, not merely unfashionable.

"Integrators look for AI." ADR-119 gave this as a reason and it does not hold: ADR-131 establishes that the section exists for editor-facing modules, so the label is read by editors. The conclusion survives its reason — "AI" is what an editor recognises, "LLM" is a developer's word — but the reason itself is withdrawn here.

Decision 

A shared top-level section netresearch_ai holds five entries:

netresearch_ai            no access check of its own
├── nrllm_aitasks         access => user     (editors)
├── nrllm_overview        access => admin    (aliases: nrllm)
├── nrllm_setup           providers, models, configurations, use-case, wizard
├── nrllm_authoring       tasks, skills, snippets
└── nrllm_operation       tools, MCP, playground, runs, analytics

Four of those five come from the fourteen entries ADR-119 called a dumping ground, which is the "three or four entries" it asks for. nrllm_aitasks is the fifth and was never one of the fourteen: it sat under web.

Consequences 

Backend shortcuts survive. They store the module identifier, and the alias resolves it. This works only because nrllm is no longer registered as a real module — an alias is shadowed by a real module of the same name.

Foreign anchors survive, and this was verified rather than assumed. t3_cowriter positions itself with ['after' => 'nrllm']. ModuleFactory::adaptAliasMappingFromModuleConfiguration() rewrites position.before and position.after through aliases, not just parent, so that anchor re-resolves to nrllm_overview with no change in the cowriter.

The container URL does not survive. /module/nrllm was the container's own path and has no successor; /module/nrllm/overview and every submodule path are unchanged. ADR-119 accepted this when it settled on keeping the submodule paths.

The three sibling extensions are not changed by this record. They may register under netresearch_ai when they choose to; nothing breaks while they do not. What this decision gives them is a place that exists.

The revisit trigger is spent. ADR-119's open question is closed, so a new editor surface is now an ordinary addition to an existing section rather than an occasion to decide where the section should be.

Alternatives 

A container under Administration, the shape the core uses for integrations. Rejected for the access-check reason above: a container inherits its parent's visibility, and Administration is admin-only, so the editor surfaces this exists for would still have had nowhere to live.

Leaving it until the trigger fired. Rejected because the trigger fires inside a change that is about something else, which is exactly how a placement gets settled by accretion. #812 existed to take the decision out of that moment.