Disciplines · Audits

Assistant terminology — one word per thing (2026-08)

BRAND.md requires the honesty disclosure to stay explicit:

5sections4 minread
Markdown source
On this page

Task 13.2 of EVE_DEEP_TEST_AND_POLISH_TODOS_2026-08-04.md. Derived from V1/BRAND.md, which is the authority; this file is the working table for the assistant surfaces and the rule each lock enforces.

The problem this closes is not untidiness. Ruling the 13.1 string inventory turned up 62 member-facing strings that call Lilith "the assistant" against 28 that call her Lilith — and the disagreement is usually inside one file, on one screen, sometimes in one sentence: indicators.ts explains persona identity twice, once as "Lilith is answering without one of her named guides" and once as "The assistant speaks generically without a named guide". A member who is introduced to Lilith and then reads about "the assistant" has been handed two names for one thing and has to work out that they are the same.

The table#

Thing Canonical, member-facing Off-terms that must not refer to it Where the other word is still correct
The assistant Lilith, or first person (I/me/my) "the assistant", "AI assistant", "Copilot", "the system", "the bot", "Eve" Operator and studio surfaces: "copilot", platform-branded (BRAND.md). The disclosure sentence below.
A guided walk of a screen tour "walkthrough", "tutorial", "spotlight tour" —
The catalogue walk audit (a run, walking flows) "sweep", "review pass", "coverage crawl" "walk" as the verb of an audit ("the audit walks every flow") is the established phrasing and stays
A captured piece of work work item "task", "ticket", "issue" as the noun for the tracked thing Builder-side only; no member surface names one. "log an issue" stays as the member's verb for creating one
A recorded choice decision "ADR" on any surface a member or operator reads ADR inside the builder graph explorer, where it is the artifact's own name
One of the six rooms room "domain" in member prose domain in ids, routes, types, and the audit catalog's structural nodes — it is the codebase's word

The one sanctioned "AI assistant"#

BRAND.md requires the honesty disclosure to stay explicit:

Responses come from Lilith, an AI assistant — not a human operator.

So "assistant" is legal as a common-noun descriptor after her name, and never as her name. "Lilith, an AI assistant" is right; "the assistant", "Open AI assistant", "Ask assistant" are not. The locks encode exactly that distinction rather than banning the word, because banning the word would delete the one sentence that has to say it.

What "member surface" means here#

Everything a member can reach: the panel and its dock, the shell chrome, the settings and profile surfaces, disclosure chips and indicators, tours, the audit board, and the invocation-point labels and accessible names that any of those render. It does not include:

  • Operator and studio surfaces, which keep platform branding and the copilot name — indicators.ts and persona-handoffs.ts carry admin entries in the same catalogues as member ones, and those are correct as they are.
  • Model-facing text: tool descriptions, input schemas, tool results, system prompt frames. The model is not a member. (It does have its own rule — the tool contract already tells it never to write a path or a domain id into a reply.)
  • Thrown errors in client libraries, because the panel renders its own copy for a failed turn rather than the exception's message.

"Eve" must appear nowhere a member can see#

Eve is the builder-side assistant. BRAND.md: never shown to members. This is checked two ways, because source and screen are different questions:

  1. A source sweep over the member-facing modules (a string can be in the tree without rendering, and a rendered string can be composed from parts).
  2. A case-insensitive sweep of the rendered text of every member route, which is the check task 13.2 asks for by name — source greps cannot see a name that arrives from an API payload or a persona record.

The word survives in code identifiers, comments, file names, the workbench and the intent ledger, all of which are builder-side and all of which BRAND.md leaves alone.

Standing item: room vs domain#

Twenty-five strings with a keep verdict still say "domain", almost all of them in the audit catalog's step actions ("Open the domain switcher", "Switch between domains", "Enter a domain"). The tour narration for the same control says "This switcher moves you between my rooms". That is a real second vocabulary and it is recorded here rather than fixed in 13.2, which is scoped to the assistant's own name: the catalog's nodes are structural (their ids are shell.navigation.domain-switch), and changing prose that quotes an id is a larger edit than this task should absorb quietly. 13.3 owns it.