Version. 0.1 Date. 2026-08-06 Status. Standing. This is the canonical statement of the rule. It replaces two independent formulations that had no single home. Author. Claude.ai (drafting session). Operator: Marvin Percival. Occasion. Operator decision, 2026-08-06, following B-11's landing. Supersedes as the authority. The rule as stated in the venture-diligence vertical lineage (from v0.2, carried unchanged to v0.4) and in the current-status manifest §5 (from v0.75, carried unchanged to v0.78). Both stand as siblings and are marked to point here.
The rule was written down twice, independently, and neither statement pointed at the other. An audit found it stated once inside one vertical's document — for a rule the other statement itself calls product-wide; every vertical inherits it — and once in a manifest entry, in different words, carried forward through four manifest versions.
Around fifteen further documents cite it, most of them pointing at a queued-directions section that only reports the rule and never originated it.
> The hazard that made this worth fixing. A correction pass that updated one lineage would leave the other live and stale, and nobody would notice, because neither names the other as authoritative. That is not hypothetical — the pass immediately before this one touched one of the two and would have missed the second.
From here: this note states the rule. Everything else cites it.
The Companion's question-answering may be demonstrated live, in front of a prospect or a customer, for an engagement holding fewer notes than the retrieval window.
Above the window, it may not.
And the universal commons is excluded regardless of its size — it exceeds the window already and grows with every account that joins, so it is never in the permitted case.
The original rule barred every live demonstration, and it was right when it was made on 2026-07-28.
Two things were wrong then. Asked the same question three times, the Companion answered correctly, then claimed it had fabricated its own correct answer, then answered correctly again — one of three states being a false statement about its own reliability. And retrieval was wholesale: the fifty most recent notes, no search of any kind, so a question about something early failed silently and looked like an absence of knowledge.
B-11 fixed the first, on 2026-08-05. Every language-model call in the engine had been deliberately pinned to answer identically except the one that writes the words a person reads. That is now pinned like the others, and two identical questions returned byte-identical answers in the live check. And an answer now carries the notes the system actually handed the model — attached by the system, not named by the model — so a reader can check the words against the material.
The second is unfixed and is why the rule only partly lifts. Retrieval is still the most recent notes up to the window, with no search. What changed is that the boundary is now visible: when an answer is partial, it says so.
And a measurement, 2026-08-06, decided where the line falls. Across every engagement holding notes, the largest anyone actually works in holds fewer than sixty, against a window of fifty. Most hold far fewer. So for a real engagement the answer is drawn from everything, and the honest scope sentence does not appear at all.
> The commons is the exception and it is not a project. It is the engagement every account joins at sign-up, it already holds more than any project, and it grows with adoption rather than with work. Demonstrating against it would demonstrate the partial case.
One thing: does the engagement hold fewer notes than the window?
If yes — the answer draws on everything, and what is shown is the product working completely.
If no — the answer is bounded, it will say so, and the rule still bars the demonstration. Honestly bounded is a harder thing to explain to a stranger than complete, and a demonstration is not the place to introduce a limit.
When the window is raised — B-12's remaining work, which needs the model's prompt capacity measured — this check gets easier and the rule does not otherwise change.
It does not bar discussing recall, in any depth, with anyone. The bar was always on demonstrating it live, and it still is above the window.
It does not bar demonstrating anything else. The rooms, the provenance walk, contribution, downloads, the record itself — none of it was ever covered by this rule and none is now.
It does not describe a defect. An answer drawn from a bounded window that says it is bounded is the product behaving correctly. The rule exists because a limit is hard to introduce in a demonstration, not because the limit is wrong.
Cite this note; do not restate the rule. Every document that needs the rule points here. A document that states it again creates a second lineage, which is the thing this note exists to end.
When the rule changes, it changes here, and the citing documents need no amendment because they carry no wording of their own.
DUNIN7 — Done In Seven LLC — Miami, Florida Loomworks — demonstrating the Companion's answers (standing note) — v0.1 — 2026-08-06 Written down twice, independently, neither pointing at the other. This is the one place.