diff --git a/docs/emergency/COLLECTIONS_SEAM.md b/docs/emergency/COLLECTIONS_SEAM.md
new file mode 100644
index 0000000..605851c
--- /dev/null
+++ b/docs/emergency/COLLECTIONS_SEAM.md
@@ -0,0 +1,143 @@
+# Collections Seam Decision
+
+## Decision
+
+The emergency fork should **reuse the collections seam**, but **not import the whole upstream collections stack as-is**.
+
+The right split is:
+
+- **reuse in place**: manifest JSON files and their resource contract
+- **reuse via adapter**: schema validation, tier resolution, and installed-status logic
+- **do not reuse directly in v1**: controller, route, and update-service wiring that is tied to Adonis, Lucid, and the current server/admin runtime
+
+In short:
+
+> the collections seam is strong enough to preserve, but the current upstream implementation is too framework-coupled to become the emergency runtime boundary unchanged.
+
+## Evidence from upstream
+
+### What is strong and worth preserving
+
+The upstream collections layer is already explicit, versioned, and content-oriented.
+
+Relevant upstream files:
+
+- `collections/kiwix-categories.json`
+- `collections/maps.json`
+- `collections/wikipedia.json`
+- `admin/types/collections.ts`
+- `admin/app/validators/curated_collections.ts`
+
+Why this is a strong seam:
+
+- manifests are versioned with `spec_version`
+- resource entries are explicit: `id`, `version`, `title`, `description`, `url`, `size_mb`
+- maps, curated ZIM content, and Wikipedia are all represented through the same family of contracts
+- validators already define the acceptable schema cleanly
+
+### What is reusable, but only behind a thin adapter
+
+Relevant upstream files:
+
+- `admin/app/services/collection_manifest_service.ts`
+- `admin/app/models/collection_manifest.ts`
+- `admin/database/migrations/1770849108030_create_create_collection_manifests_table.ts`
+- `admin/app/models/installed_resource.ts`
+- `admin/database/migrations/1770849119787_create_create_installed_resources_table.ts`
+
+Useful parts:
+
+- fetch and cache semantics for versioned specs
+- pure tier resolution logic
+- installed-status computation
+- installed resource registry shape
+
+Why it should sit behind an adapter:
+
+- `CollectionManifestService` is coupled to Axios, Vine, Lucid, and hard-coded upstream URLs
+- persistence is expressed through the current admin app models
+- emergency runtime will likely want its own daemon-owned storage and sync state
+
+### What should not be reused directly in v1
+
+Relevant upstream files:
+
+- `admin/app/controllers/easy_setup_controller.ts`
+- `admin/app/controllers/collection_updates_controller.ts`
+- `admin/start/routes.ts`
+- `admin/app/services/collection_update_service.ts`
+
+Why not:
+
+- route/controller layer is tied to the admin UX and Adonis route tree
+- update flow assumes current upstream API orchestration choices
+- emergency runtime needs `ON` / `OFF` policy enforcement and armed one-shot semantics that do not exist here
+
+## Practical reuse strategy
+
+### Reuse in place
+
+Keep these upstream assets as the canonical source-selection layer for now:
+
+- `collections/*.json`
+
+This preserves incremental compatibility and gives the emergency fork a stable seam with minimal divergence.
+
+### Reuse through an emergency adapter
+
+Create a future emergency-side adapter that owns this contract:
+
+- load a manifest spec from local file or cached remote source
+- validate against the upstream schema
+- resolve tier inheritance
+- compute installed status from local install state
+- expose normalized results to the emergency daemon
+
+This adapter should be framework-light and daemon-friendly.
+
+### Keep emergency storage independent but mappable
+
+The emergency runtime should not depend directly on Lucid models, but it should remain easy to map from upstream concepts:
+
+- `collection_manifests` -> cached upstream specs
+- `installed_resources` -> local package/install state
+
+This gives you compatibility without inheriting the current runtime stack.
+
+## Recommendation for v1
+
+For the first technical implementation:
+
+1. Treat `collections/*.json` as canonical upstream inputs.
+2. Mirror the schema contract from `admin/types/collections.ts`.
+3. Port or reimplement only the pure logic from `CollectionManifestService`:
+   - tier resolution
+   - installed-status computation
+   - filename parsing
+4. Do not reuse `CollectionUpdateService` directly.
+5. Put all remote fetch and one-shot policy logic behind the emergency daemon.
+
+## Concrete next extraction target
+
+The best next code seam is:
+
+> a pure emergency collections adapter with no Adonis dependency
+
+That adapter should accept:
+
+- manifest type
+- raw manifest JSON
+- local installed resource state
+
+And return:
+
+- validated normalized manifest
+- resolved tiers/collections
+- install status
+- candidate resources for one-shot sync
+
+## Bottom line
+
+The emergency fork should **stay aligned to upstream collections data** while **owning its own runtime behavior**.
+
+That is the cleanest way to keep incremental compatibility without dragging the entire admin/server stack into the emergency core.
diff --git a/docs/emergency/README.md b/docs/emergency/README.md
index f53573f..0388a6a 100644
--- a/docs/emergency/README.md
+++ b/docs/emergency/README.md
@@ -22,6 +22,7 @@ The profile is designed for a dedicated Android device running Ubuntu/Linux wher
 - [Architecture](./ARCHITECTURE.md)
 - [Local API](./LOCAL_API.md)
 - [Seam Map](./SEAM_MAP.md)
+- [Collections Seam](./COLLECTIONS_SEAM.md)
 - [Bootstrap Prompt](./BOOTSTRAP_PROMPT.md)
 
 ## Intended use