Version. 0.2
Date. 2026-08-05
Supersedes. v0.1, drafted and never filed. The record holds no v0.1.
Changes from v0.1. §1 is withdrawn and replaced. v0.1 raised a seed conflict that does not exist — the seed does not govern the Companion's vocabulary, and the actual authority settles the question the other way: a method fills the vocabulary wall, and the Operator governs the engagement's voice. The Operator identified this and the record already carried a prior session's identical error. What survives is smaller and real: the default method's vocabulary contains no term for three of the four rooms, which is a precondition for teaching the Companion to work in them.
Author. Claude.ai (drafting session). Operator: Marvin Percival.
Charter. standing-notes/dunin7-standing-authorization-charter-v0_1.
Build-list item. B-8 — change request D.
Grounding. inspection-briefs/loomworks-b8-step-0-findings-v0_1 — every fact about the code comes from it, read in full. standing-notes/loomworks-standing-note-skill-method-vocabulary-v0_6 — the governing authority for §1, read in full. candidate-seeds/loomworks/loomworks-candidate-seed-v0_12, read in full and found not to govern this subject.
Status. Scoping. §1 names a precondition and corrects v0.1's mistaken framing of it. The rest is options.
> v0.1 got this wrong and the correction is preserved rather than replaced. v0.1 raised a seed conflict: the seed's Voice section names Manifestation, engagement, assertion as vocabulary that appears in plain terms on operator-facing surfaces, and the Companion's persona forbids exactly those words. That framing is wrong, and the record already contained its refutation.
>
> The seed does not govern the Companion's vocabulary. It mentions Companion once, as a methodology noun, and has no Companion-configuration section at all. A prior session made this same error — claiming seed authority for a Companion-customization model that does not exist there — and the correction is filed at investigations/ as the Companion Skill Customization Verification Response. The drafting session did not check the record before raising a conflict, which is the failure this project keeps correcting.
The actual authority is standing-notes/loomworks-standing-note-skill-method-vocabulary-v0_6, with manifest §672. And it settles the question in the opposite direction from v0.1's reading.
A method fills the vocabulary wall. §6.2 of that note says it directly, with worked examples: the coaching method makes "session / client / intake" the surface words; the litigation method makes "matter / privilege / production" the surface words. The wall is a slot a method fills, not a rule the engine imposes — and §7 settles that the engagement's voice is the Operator's to govern, singular, never an automatic average.
So the Companion's language is a product variable, by settlement. It is not forced to be the engine's language, it is customizable per method and per vertical, and the persona's word list is the default method's fill — a choice, not a violation. The seed's Voice section governs written operator-facing surfaces; the standing note's §8 separates the interaction surface the installed method drives from the production surface Shaping drives. Both are true because they govern different surfaces.
The default fill has no word for three of the four rooms.
The persona forbids manifestation, shape and render and offers project, note, specification, artifact, draft, saved, waiting in their place. Nothing in that replacement set names the Manifestation room, the Shaping room, or the Rendering room. It has words for things — a note, an artifact, a draft — and none for the places.
That is not a wrong word. It is no word. A Companion running the default method cannot refer to those three rooms in any register, including a plain-English one. Teaching it to work in three rooms it cannot name is the precondition, and it is small: the default method's fill needs terms for them.
What the terms should be is the Operator's, per §7 — the engagement speaks in one voice and the Operator governs it. They need not be the room names. Your summary, the arrangement for a reader, the finished piece would all serve. What they cannot be is absent.
Four rooms, four different situations, and only one of them is finished.
| Room | Read | Write | |---|---|---| | Memory | Full | Full | | Manifestation | None at all | None | | Shaping | Counts only — how many are waiting, never what they are | None that works | | Rendering | A narrow list of finished artifacts, plus counts | None that works |
Two details worth knowing beyond the table.
Shaping and Rendering's writes are not merely absent — they are wired to a mechanism that dead-ends. A request to produce a draft routes through the delegation and authorization layer, which is real, default-deny, and works — and then finds no handler registered for either capability. The gate would bind the moment a handler appeared; there is simply nothing behind it.
And the Companion cannot be tricked into a write by a vague sentence. A request that classifies poorly collapses to ordinary conversation, and that path exits before any write handler is entered. It is not a prompt restraint — the code is never reached. (The build list has carried this as an open question since it was written. It is closed, in the negative.)
> But the safety is incidental, and the findings say so plainly. The wiring for the sharpest case was never built — which is a different finding than the wiring being safe. Anything B-8 builds has to hold that property deliberately, because today it holds by absence.
Mostly wiring, with one exception and one genuine gap.
Reads across all three rooms are wiring. The engine already serves everything — list and get for manifestations, shape events, considerations, renders. The Companion would call them the way it already calls Memory's. This is the same finding every room's build produced: the engine serves what the surface needs.
One thing is missing rather than unwired. No handler is registered for initiating a render, so even a fully authorized request has nothing to call. That is a build, not a wire.
And one thing should stay where it is. Confirming a Shape enqueues production of renders elsewhere, and the Shaping room's own build made its control name what it will produce before it is pressed. A sentence cannot carry that property. Confirm the thing and confirm the thing, and here is what will be made are different acts, and only the second is the Operator approving something. This one belongs on a surface with a preview, not in a conversation.
4.1 Sources do not generalize. B-11 made an answer's citations facts rather than model claims — but only for one intent, hard-coupled to one key in one dictionary. Four other retrieval-backed answers already produce no sources today. An answer about Manifestation or Rendering would carry none unless separately wired.
That is a decision, not an oversight to fix in passing. Either sources become general — the system attaches what it retrieved, whatever the intent — or answers about the other rooms are less checkable than answers about Memory, and a person cannot tell which kind they are looking at.
4.2 The membership fence is one check at one place. Everything the Companion reaches sits behind a single gate at the conversation's entry point, and the room endpoints' own gates are not in its path — it queries beneath them. A new capability either stays behind that one gate or calls the rooms' HTTP routes and inherits theirs. Both work; mixing them silently is how a gap appears.
Reads first, writes second, and they are genuinely separable.
Slice one — the Companion can see and describe all four rooms. Reads only. No new authority question, because reading changes nothing. This is the slice that makes the Companion useful across the whole system, and it is mostly wiring.
Slice two — the writes worth having, chosen deliberately. Deriving a Manifestation and asking for a draft are the plausible ones. Each needs its authority answered on its own terms, not inherited from Memory's.
And confirming a Shape is in neither slice. §3.
§1 first, and the choice of words is yours. A Companion built to work in three rooms it has no word for cannot discuss them. The terms need not be the room names — the standing note settles that the Operator governs the engagement's voice — but they have to exist.
Then slice one. It is wiring, it changes nothing, and it turns a Companion that knows a quarter of the system into one that knows all of it.
Then the sources decision at §4.1, before slice two rather than after — because retrofitting a citation channel onto answers already shipped without one means changing what a person has learned to expect.
Slice two last, per operation, with its authority stated rather than inherited.
DUNIN7 — Done In Seven LLC — Miami, Florida Loomworks — B-8 scoping note — v0.2 — 2026-08-05 The Companion has no word for three of the four rooms. Not the wrong word — no word.