diff --git a/docs/campaign-repositories.md b/docs/campaign-repositories.md
new file mode 100644
index 0000000..4705b12
--- /dev/null
+++ b/docs/campaign-repositories.md
@@ -0,0 +1,141 @@
+# The `pr-` convention — campaign repositories
+
+A `pr-` repository is a **campaign**: an ongoing public-communication offering
+with its own voice, cadence, consent records and budget. It consumes one or more
+`fluid-*` interfaces and owns none of them.
+
+This document exists so the second campaign repository does not have to be
+reverse-engineered from the first. `pr-hall-of-helix` is the first instance.
+
+## Why the prefix says `pr-`
+
+The prefix covers public relations, promotion, **and propaganda** — including the
+uncomfortable third word deliberately. A campaign repository that cannot name
+what it is doing will not govern it well.
+
+The distinction between the three is not the mechanism, which is identical in all
+cases. It is whether the campaign is accountable to something outside itself: a
+source document it must not contradict, a named reviewer, a consent basis it can
+lose. Everything in this convention exists to keep that accountability
+structural rather than cultural, because culture does not survive a deadline.
+
+## What makes a repository a campaign
+
+Three properties, together. A repository with only one or two of them is
+something else.
+
+1. **It speaks outward, in an organisation's voice**, to an audience that did not
+   ask to be written to.
+2. **It writes about people** — usually people who did not commission the writing.
+3. **It does not finish.** A campaign has a cadence, not a completion criterion.
+   This is why campaigns must not live inside a workplan: a workplan that never
+   finishes is a workplan that has stopped being honest.
+
+## Required contents
+
+### `INTENT.md`
+
+The ordinary repository intent, plus the campaign's own three questions
+answered explicitly:
+
+- **Audience** — who is being written to, and what they did to deserve the
+  attention (usually nothing; they are strangers, and that is the point).
+- **Voice** — how the organisation sounds here, and why that is appropriate.
+- **What it will not say.** The most important of the three and the one most
+  often skipped. A campaign with no stated refusals has not decided anything; it
+  has merely not yet been tempted.
+
+### Consent records
+
+For every person the campaign writes about: the basis, its date, and how it may
+be revoked.
+
+Three rules that are not the campaign's to relax:
+
+- **Absence of objection is not consent.** Silence is not agreement, and treating
+  it as such is the single most common failure in this class of work.
+- **A standing agreement does not apply retroactively.** People whose
+  contributions predate the agreement are asked, not assumed in.
+- **Revocation is honoured going forward**, and the record says so before anyone
+  needs it.
+
+### Composition strategies as named variants
+
+Every composed artifact records the `variant` label of the strategy that produced
+it, at composition time.
+
+Not afterwards. Reconstructing which strategy wrote what from timestamps is the
+kind of inference that works until it quietly stops, and by then the comparison
+it feeds has already been believed.
+
+### A publishing cadence
+
+Stated, and slow enough to review at. A backlog is a source to draw on, not a
+queue to drain — publishing it in a burst is a rate-limit problem, an unkind
+introduction for a new subscriber, and a volume of writing nobody has had time to
+read properly.
+
+### A review step
+
+A person reads each artifact against its source before it goes out, and their
+name travels with it.
+
+This is not ceremony. A condensation can put a claim in someone's mouth that
+their own source never made, and **in an audit trail an unfaithful artifact looks
+exactly like a faithful one.** No downstream check recovers this, because there
+is nothing anomalous to detect. The reviewer is the only mechanism that works.
+
+## The boundary against `fluid-*`
+
+| Campaign (`pr-`) owns | Interface (`fluid-*`) owns |
+|---|---|
+| Composition and the editorial voice | The contract and its revisions |
+| Choosing what is worth publishing | The adapter: rendering and delivery |
+| Variant strategies and their competition | Delivery experiments |
+| Consent records and the review queue | Per-variant engagement reporting |
+| The publishing cadence | Platform and channel configuration |
+| The *declared* presence (names, descriptions, assets) | *Applying* that declaration, and custody of every credential |
+
+A campaign reaches its interface through the published contract and nothing else.
+It does not import the adapter, read the evidence store, or hold a credential.
+
+The separation is architectural rather than stylistic: composition is a
+probabilistic editorial task, and `ArchitectureBlueprint.md` §48.1 forbids an LLM
+in the mandatory request path. Putting composition in the campaign is what keeps
+the interface deterministic.
+
+### Presence lives with the campaign
+
+A bot's name, its description, and the channel titles a stranger reads are
+**brand, not mechanism**, so the declaration belongs to the campaign. The
+interface holds the schema, applies the declaration, and owns the credentials
+that result. See `fluid-telegram/docs/provisioning.md` for the worked case.
+
+## Registration
+
+- Repository name: `pr-<campaign-slug>`
+- State Hub domain: **the domain of the subject being promoted**, not a domain
+  named for marketing. The domains are market sectors, and a campaign is
+  classified by what it is about. `pr-hall-of-helix` is `infotech` because
+  HelixForge is.
+- Topic: the subject's topic, so a query for that subject's work returns the
+  campaign with it (`pr-hall-of-helix` → `helix-forge`).
+- Workplan prefix: derived from the repo slug (`PRHOH-WP-`).
+
+## Its own FLUID loop
+
+A campaign is a consumer of an interface, and it may also run its own FLUID loop
+over its own artifacts — which composition variant travels — using the
+per-variant engagement the interface reports.
+
+These are different loops over different questions and should not be merged. The
+interface's loop asks *how should this be delivered*; the campaign's asks *what
+should be said*. The first is settled by the platform's constraints; the second
+never fully settles.
+
+## The test of this convention
+
+Whether a second campaign can be created from this document without reading
+`pr-hall-of-helix`. Until that has happened, treat the convention as a
+description of one case rather than a general rule — see `FLUID-WP-0009` T05,
+which makes the same point about the framework itself.
diff --git a/workplans/FLUID-WP-0009-campaign-repos-and-example-separation.md b/workplans/FLUID-WP-0009-campaign-repos-and-example-separation.md
index 037bb7a..503c9b8 100644
--- a/workplans/FLUID-WP-0009-campaign-repos-and-example-separation.md
+++ b/workplans/FLUID-WP-0009-campaign-repos-and-example-separation.md
@@ -46,7 +46,7 @@ the wrong place.
 
 ```task
 id: FLUID-WP-0009-T01
-status: todo
+status: done
 priority: medium
 state_hub_task_id: "2a3aab85-e464-5b17-980f-1896ecd3794f"
 ```
@@ -63,6 +63,16 @@ audience, voice and what it will not say; consent records for anyone it writes
 about; composition strategies as named variants; and a publishing cadence. It
 consumes one or more `fluid-*` interfaces and owns none of them.
 
+**Written: `docs/campaign-repositories.md` (2026-09-04.)** It adds a review step
+to the list above, the boundary table against `fluid-*`, the registration rules
+(domain and topic follow the *subject*, since the domains are market sectors and
+none of them is "marketing"), and the point that a campaign may run its own FLUID
+loop over its own artifacts without merging it into the interface's.
+
+It states its own test: whether a second campaign can be created from the
+document without reading `pr-hall-of-helix`. Until that happens it describes one
+case rather than a general rule — the same caution T05 applies to the framework.
+
 ## T02 — Establish pr-hall-of-helix
 
 ```task