Version. 0.1
Date. 2026-08-05
Author. Claude.ai (drafting session). Operator: Marvin Percival.
Charter. standing-notes/dunin7-standing-authorization-charter-v0_1.
Build-list item. B-8, slice two. Slice one closed at CR-2026-170.
Grounding. inspection-briefs/loomworks-b8-step-0-findings-v0_1; scoping-notes/loomworks-b8-scoping-note-v0_2; candidate-seeds/loomworks/loomworks-candidate-seed-v0_12 for the authority principle; change-requests/cr-2026-163-shaping-room-v0_2 and cr-2026-170-companion-room-reads-v0_1 for two decisions this note reopens.
Status. Scoping. §3 reopens a decision taken twice and puts it to the Operator. §5 is a decision that should come before any build.
The seed is explicit: automatic state transitions on artifacts the Operator has authority over are a category error. The system surfaces and signals; the Operator approves.
A conversational write sits awkwardly against that, and the awkwardness is the whole subject. When the Operator types make me a draft of the onboarding guide, is that the Operator approving, or the system acting on an interpretation?
It is both, and which one dominates depends on a single property: whether the Operator can know what they are approving from the sentence they typed.
> The Shaping room's build already established this, in a different form. Its confirm control names which outputs it will produce before it is pressed, because approval requires knowing what is being approved — a button that silently enqueues several renders is the Operator's authority in name and not in substance. > > The same test applies to a sentence, and it is what separates the writes below from each other.
So each candidate write is examined on one question: is its consequence knowable from the request?
Knowable, bounded, cheap. Deriving a Manifestation. The Operator asks for their notes to be organised; the system organises them. The consequence is one new reading of the record, and the prior one is preserved rather than replaced. Nothing is spent, nothing is destroyed, and the result is visible immediately. The sentence and the act are the same size.
Knowable in kind, not in extent. Asking for a draft. Produce a draft of X names the artifact. What it does not name is the cost — it spends against the engagement's credit, and the Operator typing the sentence does not know how much. The consequence is one artifact, which is knowable; the price is not.
Not knowable from the request at all. Confirming a Shape. It enqueues production of every render type the engagement has declared for that shape. **The Operator saying confirm it cannot know how many outputs that produces or what they are** — which is exactly why the room's control shows them first.
Excluded from this slice regardless of authority. Anything in Memory. Already fully available and unchanged.
CR-2026-163 put the confirm on a surface. CR-2026-170 said it stays there in both slices. Both on the same reasoning: a sentence cannot carry the property that a control carries — confirm the thing and confirm the thing, and here is what will be made are different acts.
That reasoning was about a sentence. It is not true of an exchange.
> The Companion could state the consequence and wait. Confirming this will produce a PDF and an HTML version of the care guide. Say go and I will. The Operator then knows exactly what they are approving, and says so. That is the control's property — consequence named before commitment — expressed in the medium the Companion actually has. > > I did not consider this when I wrote either decision, and the argument I gave in both was narrower than I made it sound.
What would have to be true for it to be sound, and none of it is established:
The consequence must be computed, not described. The Shaping room computes the list from the engagement's declared render types and the mapping selector. A Companion saying what will be produced must call the same computation — not ask a model to summarise what it thinks will happen. If the model composes that sentence, the Operator is approving a claim rather than a fact, which is the defect B-11 removed from citations.
The confirmation must be unambiguous and bound to that statement. Go after a list is a different thing from go in a stream of conversation. What ties one to the other, and what happens if the Operator says something ambiguous, is unread.
And the room's own control must remain the primary path. A conversational route that becomes the usual way to confirm would move an approval off a surface that shows a preview onto one that does not.
This is the Operator's. Three positions are available: the confirm stays surface-only as decided twice; it becomes reachable by an explicit two-turn exchange under the conditions above; or the question waits until the other two writes have been built and used.
4.1 What the authorization layer actually requires. A request to produce a draft routes through a delegation and authorization layer that is real, default-deny, and working — and finds no handler registered. What that layer would demand of a registered handler is unread: a capability grant, an approval record, a credit check, or something else. It may already implement most of what §3 would need.
4.2 What deriving a Manifestation costs. Named here as cheap on the reasoning that it reorganises existing records. That is inference, not a read. If it invokes a model, it is not cheap and §2's first row moves.
4.3 Whether an approval record exists. The seed's the Operator approves implies something recorded, not merely something done. Whether the engine has such a thing — and whether a conversational approval could produce one — is unread.
4.4 What the credit spend actually is, and whether it can be stated in advance. §2's second row turns on this. If the cost of a draft is knowable before it is produced, the Companion can say it and the write becomes knowable. If it is not, the Operator is approving an unbounded amount.
Sources still do not generalize. An answer's citations are facts for one intent, hard-coupled to one key; slice one's three new answers carry none, and the implementation notes record that.
This matters more for writes than for reads. A read that cannot show its sources is less useful. A write proposed on the basis of material the Companion cannot show is an approval the Operator cannot check — I'll draft this from your notes about X is not verifiable unless the notes are attached.
Settle it before slice two rather than after. Retrofitting a citation channel onto a write flow already shipped means changing what the Operator has learned to expect from an approval.
Build the cheapest knowable write first, alone: deriving a Manifestation — conditional on §4.2 confirming it is cheap. It is the only candidate whose consequence and cost are both knowable from the request as typed, and it exercises the whole path without any authority question the seed has not already answered.
Then settle §5, then the draft write with its cost stated in advance if §4.4 allows and not at all if it does not.
And leave §3 until both have been used. The two-turn confirmation is a real mechanism and I am not confident enough in it to recommend it having twice argued the opposite — and the writes below it will teach more about whether conversation can carry an approval than any amount of reasoning about it now.
DUNIN7 — Done In Seven LLC — Miami, Florida Loomworks — B-8 slice two scoping note — v0.1 — 2026-08-05 The test is whether the Operator can know what they are approving from the sentence they typed.