The question this folds into. Document management systems — Google's ecosystem and its peers — hold most of what an organization already knows. Can Loomworks source Memory material from them? The June arc answered yes, with one rule that carries everything: the source supplies; the Operator governs. Material can flow in from any managed store, but it enters Memory carrying provenance assigned at the boundary — where it came from, what kind of origin it has — and nothing about the source system's own confidence survives unexamined.
The finding that matters most, still. Google's Open Knowledge Format captures knowledge cheaply — an enrichment agent crawls an estate and emits a confident description of every concept — and carries no origin or trust field. The legacy-estate use case named the consequence: cheap capture makes a model's inference, a retiring engineer's recollection, and a possibly-wrong 1987 design note look equally authoritative. That is precisely the failure Loomworks's provenance threads exist to prevent, and it converts the document-management wave from a threat into a feedstock argument: the more knowledge these systems capture flatly, the more valuable the governed record that receives it with origin restored.
What this note does. Places each June finding into the thread it now belongs to, names the corrections the arc needs (one vocabulary, one verification), and asks three decisions.
Four documents, one line each, per the session record:
Folding the arc's rule into today's architecture, sourcing from a document management system has three parts, all of which now have homes:
The channel is an ingestion adapter — the engagement-as-universal-adapter pattern, on the ingestion-interfaces arc where the spreadsheet and presentation work already sits. A document-management source (a Drive folder, a SharePoint library, an OKF-emitting repository) is a source class on that arc, not a new architecture. The Phase 57-era external-interfaces scoping already framed the trigger shapes (pull-at-read, scheduled pull, push, watch); a managed store is the watch-or-scheduled-pull case.
The boundary is where provenance is assigned, and it is the whole game. Material entering from an external store gets outward provenance threads at ingestion: the source system, the document, the capture time — and, for OKF-shaped material specifically, an origin classification the format itself does not carry. Machine-enriched concepts enter machine-origin-marked, non-suppressibly, exactly as any machine-assisted contribution does. The trust axis is tie-to-verifiable-source; a confidently-worded concept file with no traceable origin enters as exactly that, visibly. The June hazard — flat authority — is answered structurally, not editorially.
The governance is unchanged: source-supplies, Operator-governs. Held-until-admitted applies to external-sourced material the way it applies to outside contributors; the store does not write into committed Memory because it exists — the team admits what enters the record.
| June finding | Current home |
|---|---|
| Document stores as Memory sources | The ingestion-interfaces arc: a document-management source-adapter class, joining the spreadsheet/presentation work already queued |
| OKF consume (and possibly emit) | Standing candidate direction — unchanged status, honesty flag intact; a format the adapter class could speak, decided when a consumer exists |
| Flat-authority hazard | The boundary rule in §2 — provenance-at-ingestion; already consistent with the provenance-threads direction queued for seed v0.13 |
| Corpus effect | ask-your-engagement: store-sourced document text widens the askable corpus, and answer citations must carry the external source identity — a requirement that arc's scoping note should carry explicitly |
| Two sourcing shapes | The Hebbia comparison's instrument-and-record composition gains a sibling: parse the repository (Hebbia's shape) and consume the emitted format (OKF's shape) are both adapter work, and the record governs both identically |
| Positioning playbook's eight value propositions | Input to the marketing-site copy draft when D-5 commissions it |
| Non-transitivity / composition-in-Shaping correction | Already aligned with the current seed's framing; noted here so the fold-in does not re-litigate it |
Two, neither urgent, both recorded so they are not lost:
Vocabulary. The June documents cite "OVA authorisation reach." The GRANTHA standing correction (2026-07-28) applies: GRANTHA holds the authorization seat, grant as sole primitive. The arc's documents are historical record and are not rewritten; the correction applies at their next version bump, and any outward reuse of the playbook substitutes the current seat.
Verification. The 2026-07-06 session recorded four open items from the arc that this note cannot resolve from here: the branch merge status (whether add-okf-loomworks-document-arc-v0_1 ever merged — if not, the four pairs are not on the record's main and filing them is overdue); the restatement use case's v0.2 bump decision; an untracked companion functional spec v0.3 against a canonical v0.1; and the OKF interoperability scoping item itself. The Claude Code retrieval below reports all four.
D-1 — The source-adapter class. Add "document-management sources" as a named class on the ingestion-interfaces arc (Drive/SharePoint-style stores; OKF as a candidate format within it), scoped when that arc next moves. Recommendation: yes — it is a scoping line-item today, not a build.
D-2 — OKF's status. Reaffirm consume-and-emit as candidate direction, honesty flag standing — or promote it. Recommendation: reaffirm; no consumer is pressing, and the flag has kept outward material honest.
D-3 — The arc's housekeeping. On the retrieval's findings: merge or file the four pairs if the branch never landed, and take the restatement v0.2 decision then. Recommendation: resolve with one CC pass when the retrieval reports.