diff --git a/labeling/backend/README.md b/labeling/backend/README.md
index d659d75..2152ed2 100644
--- a/labeling/backend/README.md
+++ b/labeling/backend/README.md
@@ -4,13 +4,10 @@ API + data model for the bounding-box labeling tool. See CLAUDE.md §3.2 for the
 full requirements this is built against.
 
 **Status:** core data model + CRUD API implemented (sets, hierarchical classes,
-frames, labels, per-frame/per-set status), frame image serving, and bulk frame
-ingest from a recording session (see `labeling/frontend/` for the UI). **Not yet
-implemented:** dataset versioning/promotion (`enemy-v1`, `enemy-v2`, ...) —
-deferred since it needs its own design pass (snapshot semantics: labels are
-editable at any time, but a promoted dataset version must stay reproducible).
-Also not yet implemented: auth/login (see below) or the active-learning
-auto-label workflow.
+frames, labels, per-frame/per-set status), frame image serving, bulk frame
+ingest from a recording session, and dataset version promotion (see
+`labeling/frontend/` for the UI). **Not yet implemented:** auth/login (see
+below) or the active-learning auto-label workflow.
 
 ## Stack
 
@@ -32,6 +29,11 @@ auto-label workflow.
   frame + set + sub-class.
 - `FrameSetStatus` — per-frame, per-set label state (`unlabeled` / `auto_labeled` /
   `reviewed`).
+- `DatasetVersion` / `DatasetVersionFrame` / `DatasetVersionLabel` — a named,
+  immutable snapshot of a set's frames+labels as of promotion time (`enemy-v1`,
+  `enemy-v2`, ...). Labels stay live-editable at any time; a promoted version
+  copies the label data at that moment, so it stays reproducible regardless of
+  later edits or deletions to the live labels.
 
 Frame images are served from `FRAMES_ROOT` (env var, default `./frames`) at
 `/images/<image_path>`; `Frame.image_path` is always relative to that root.
@@ -71,6 +73,23 @@ Safe to re-run: frames already registered for that session (by frame index) are
 skipped. `--width`/`--height` default to 1280x720 (Spelunky Classic HD's fixed
 capture resolution).
 
+## Promoting a dataset version
+
+Once frames are marked `reviewed` (see the frontend, or `PUT
+/frames/{id}/sets/{id}/status`), freeze them into a named, reproducible
+version:
+
+```
+POST /sets/{set_id}/dataset-versions
+{"name": "enemy-v1", "description": "first reviewed batch"}
+```
+
+Defaults to every currently-`reviewed` frame in that set; pass an explicit
+`frame_ids` list to promote a different selection instead. Fetch the frozen
+result (what training will eventually consume) via `GET
+/dataset-versions/{id}` — it returns each frame plus the exact labels that
+existed at promotion time, unaffected by any later edits to the live labels.
+
 ## Testing
 
 ```
diff --git a/labeling/backend/src/spelunkai_labeling_backend/main.py b/labeling/backend/src/spelunkai_labeling_backend/main.py
index 3d77d29..aea78b0 100644
--- a/labeling/backend/src/spelunkai_labeling_backend/main.py
+++ b/labeling/backend/src/spelunkai_labeling_backend/main.py
@@ -17,6 +17,7 @@ from fastapi.middleware.cors import CORSMiddleware
 from fastapi.staticfiles import StaticFiles
 
 from .db import Database, get_db
+from .routers.dataset_versions import router as dataset_versions_router
 from .routers.frames import router as frames_router
 from .routers.labels import router as labels_router
 from .routers.sets import main_classes_router, sets_router
@@ -49,6 +50,7 @@ def create_app(database_url: Optional[str] = None, frames_root: Optional[str] =
     app.include_router(main_classes_router)
     app.include_router(frames_router)
     app.include_router(labels_router)
+    app.include_router(dataset_versions_router)
 
     app.mount("/images", StaticFiles(directory=frames_dir), name="images")
 
diff --git a/labeling/backend/src/spelunkai_labeling_backend/models.py b/labeling/backend/src/spelunkai_labeling_backend/models.py
index 692c86c..e65b3a3 100644
--- a/labeling/backend/src/spelunkai_labeling_backend/models.py
+++ b/labeling/backend/src/spelunkai_labeling_backend/models.py
@@ -119,3 +119,46 @@ class Label(Base):
     updated_at = Column(DateTime, default=_utcnow, onupdate=_utcnow)
 
     sub_class = relationship("SubClass")
+
+
+class DatasetVersion(Base):
+    """A named, immutable snapshot of a set's reviewed frames+labels, for
+    reproducible training runs (labels stay editable at any time, but a
+    promoted version's contents never change)."""
+
+    __tablename__ = "dataset_versions"
+    __table_args__ = (UniqueConstraint("set_id", "name", name="uq_dataset_version_set_name"),)
+
+    id = Column(Integer, primary_key=True)
+    set_id = Column(Integer, ForeignKey("label_sets.id"), nullable=False)
+    name = Column(String, nullable=False)
+    description = Column(String, nullable=True)
+    created_at = Column(DateTime, default=_utcnow)
+    created_by_id = Column(Integer, ForeignKey("users.id"), nullable=True)
+
+
+class DatasetVersionFrame(Base):
+    """Which frames are included in a dataset version."""
+
+    __tablename__ = "dataset_version_frames"
+    __table_args__ = (UniqueConstraint("dataset_version_id", "frame_id", name="uq_dataset_version_frame"),)
+
+    id = Column(Integer, primary_key=True)
+    dataset_version_id = Column(Integer, ForeignKey("dataset_versions.id"), nullable=False)
+    frame_id = Column(Integer, ForeignKey("frames.id"), nullable=False)
+
+
+class DatasetVersionLabel(Base):
+    """A frozen copy of a Label as of promotion time - the live Label row it
+    was copied from may later be edited or deleted without affecting this."""
+
+    __tablename__ = "dataset_version_labels"
+
+    id = Column(Integer, primary_key=True)
+    dataset_version_id = Column(Integer, ForeignKey("dataset_versions.id"), nullable=False)
+    frame_id = Column(Integer, ForeignKey("frames.id"), nullable=False)
+    sub_class_id = Column(Integer, ForeignKey("sub_classes.id"), nullable=False)
+    x = Column(Float, nullable=False)
+    y = Column(Float, nullable=False)
+    width = Column(Float, nullable=False)
+    height = Column(Float, nullable=False)
diff --git a/labeling/backend/src/spelunkai_labeling_backend/routers/dataset_versions.py b/labeling/backend/src/spelunkai_labeling_backend/routers/dataset_versions.py
new file mode 100644
index 0000000..024eaf1
--- /dev/null
+++ b/labeling/backend/src/spelunkai_labeling_backend/routers/dataset_versions.py
@@ -0,0 +1,129 @@
+"""Dataset version promotion: frozen, reproducible snapshots of a set's
+reviewed frames and labels, for training to consume without being affected
+by later labeling edits.
+"""
+from __future__ import annotations
+
+from fastapi import APIRouter, Depends, HTTPException
+from sqlalchemy.orm import Session
+
+from .. import models, schemas
+from ..db import get_db
+from ..users import get_or_create_user
+
+router = APIRouter(tags=["dataset-versions"])
+
+
+@router.post("/sets/{set_id}/dataset-versions", response_model=schemas.DatasetVersionSummary, status_code=201)
+def create_dataset_version(set_id: int, payload: schemas.DatasetVersionCreate, db: Session = Depends(get_db)):
+    if db.get(models.LabelSet, set_id) is None:
+        raise HTTPException(404, "set not found")
+    if db.query(models.DatasetVersion).filter_by(set_id=set_id, name=payload.name).first():
+        raise HTTPException(409, f"dataset version '{payload.name}' already exists for this set")
+
+    if payload.frame_ids is not None:
+        frame_ids = payload.frame_ids
+        found = {
+            row.id for row in db.query(models.Frame.id).filter(models.Frame.id.in_(frame_ids))
+        }
+        missing = sorted(set(frame_ids) - found)
+        if missing:
+            raise HTTPException(422, f"unknown frame ids: {missing}")
+    else:
+        frame_ids = [
+            row.frame_id
+            for row in db.query(models.FrameSetStatus.frame_id).filter_by(
+                set_id=set_id, status=models.LabelStatus.REVIEWED
+            )
+        ]
+    if not frame_ids:
+        raise HTTPException(422, "no frames to promote (no reviewed frames, and no frame_ids given)")
+
+    version = models.DatasetVersion(
+        set_id=set_id,
+        name=payload.name,
+        description=payload.description,
+        created_by_id=get_or_create_user(db, payload.created_by).id if payload.created_by else None,
+    )
+    db.add(version)
+    db.flush()  # assigns version.id without committing yet
+
+    for frame_id in frame_ids:
+        db.add(models.DatasetVersionFrame(dataset_version_id=version.id, frame_id=frame_id))
+        for label in db.query(models.Label).filter_by(frame_id=frame_id, set_id=set_id).all():
+            db.add(models.DatasetVersionLabel(
+                dataset_version_id=version.id,
+                frame_id=frame_id,
+                sub_class_id=label.sub_class_id,
+                x=label.x,
+                y=label.y,
+                width=label.width,
+                height=label.height,
+            ))
+
+    db.commit()
+    return schemas.DatasetVersionSummary(
+        id=version.id,
+        set_id=version.set_id,
+        name=version.name,
+        description=version.description,
+        frame_count=len(frame_ids),
+    )
+
+
+@router.get("/sets/{set_id}/dataset-versions", response_model=list[schemas.DatasetVersionSummary])
+def list_dataset_versions(set_id: int, db: Session = Depends(get_db)):
+    if db.get(models.LabelSet, set_id) is None:
+        raise HTTPException(404, "set not found")
+
+    results = []
+    for version in db.query(models.DatasetVersion).filter_by(set_id=set_id).all():
+        frame_count = db.query(models.DatasetVersionFrame).filter_by(dataset_version_id=version.id).count()
+        results.append(schemas.DatasetVersionSummary(
+            id=version.id,
+            set_id=version.set_id,
+            name=version.name,
+            description=version.description,
+            frame_count=frame_count,
+        ))
+    return results
+
+
+@router.get("/dataset-versions/{version_id}", response_model=schemas.DatasetVersionDetail)
+def get_dataset_version(version_id: int, db: Session = Depends(get_db)):
+    version = db.get(models.DatasetVersion, version_id)
+    if version is None:
+        raise HTTPException(404, "dataset version not found")
+
+    frame_ids = [
+        row.frame_id
+        for row in db.query(models.DatasetVersionFrame.frame_id).filter_by(dataset_version_id=version_id)
+    ]
+    frames_by_id = {
+        f.id: f for f in db.query(models.Frame).filter(models.Frame.id.in_(frame_ids))
+    } if frame_ids else {}
+
+    labels_by_frame: dict[int, list[models.DatasetVersionLabel]] = {}
+    for label in db.query(models.DatasetVersionLabel).filter_by(dataset_version_id=version_id).all():
+        labels_by_frame.setdefault(label.frame_id, []).append(label)
+
+    frames_out = [
+        schemas.DatasetVersionFrameRead(
+            frame=schemas.FrameRead.model_validate(frames_by_id[frame_id]),
+            labels=[
+                schemas.DatasetVersionLabelRead(
+                    sub_class_id=label.sub_class_id, x=label.x, y=label.y, width=label.width, height=label.height,
+                )
+                for label in labels_by_frame.get(frame_id, [])
+            ],
+        )
+        for frame_id in frame_ids
+    ]
+
+    return schemas.DatasetVersionDetail(
+        id=version.id,
+        set_id=version.set_id,
+        name=version.name,
+        description=version.description,
+        frames=frames_out,
+    )
diff --git a/labeling/backend/src/spelunkai_labeling_backend/schemas.py b/labeling/backend/src/spelunkai_labeling_backend/schemas.py
index 166f2d5..8d2fd69 100644
--- a/labeling/backend/src/spelunkai_labeling_backend/schemas.py
+++ b/labeling/backend/src/spelunkai_labeling_backend/schemas.py
@@ -112,3 +112,41 @@ class FrameSetStatusUpdate(BaseModel):
 class FrameWithStatus(BaseModel):
     frame: FrameRead
     status: LabelStatus
+
+
+class DatasetVersionCreate(BaseModel):
+    name: str
+    description: Optional[str] = None
+    created_by: Optional[str] = None
+    # Which frames to freeze into the version. Default (None): every frame
+    # currently `reviewed` for this set - the normal "promote what's done" case.
+    frame_ids: Optional[list[int]] = None
+
+
+class DatasetVersionSummary(BaseModel):
+    id: int
+    set_id: int
+    name: str
+    description: Optional[str] = None
+    frame_count: int
+
+
+class DatasetVersionLabelRead(BaseModel):
+    sub_class_id: int
+    x: float
+    y: float
+    width: float
+    height: float
+
+
+class DatasetVersionFrameRead(BaseModel):
+    frame: FrameRead
+    labels: list[DatasetVersionLabelRead]
+
+
+class DatasetVersionDetail(BaseModel):
+    id: int
+    set_id: int
+    name: str
+    description: Optional[str] = None
+    frames: list[DatasetVersionFrameRead]
diff --git a/labeling/backend/tests/test_dataset_versions.py b/labeling/backend/tests/test_dataset_versions.py
new file mode 100644
index 0000000..56f882b
--- /dev/null
+++ b/labeling/backend/tests/test_dataset_versions.py
@@ -0,0 +1,111 @@
+def _make_set_with_subclass(client, name="Enemy", sub_name="Bat"):
+    set_id = client.post("/sets", json={"name": name}).json()["id"]
+    main_class_id = client.post(f"/sets/{set_id}/main-classes", json={"name": name}).json()["id"]
+    sub_class_id = client.post(f"/main-classes/{main_class_id}/sub-classes", json={"name": sub_name}).json()["id"]
+    return set_id, sub_class_id
+
+
+def _make_frame(client, session_name="run01", frame_index=0):
+    payload = {"session_name": session_name, "frame_index": frame_index, "image_path": "x.png", "width": 1280, "height": 720}
+    return client.post("/frames", json=payload).json()["id"]
+
+
+def _make_reviewed_frame_with_label(client, set_id, sub_class_id, frame_index):
+    frame_id = _make_frame(client, frame_index=frame_index)
+    client.post(
+        f"/frames/{frame_id}/sets/{set_id}/labels",
+        json={"sub_class_id": sub_class_id, "x": 1, "y": 2, "width": 3, "height": 4},
+    )
+    client.put(f"/frames/{frame_id}/sets/{set_id}/status", json={"status": "reviewed"})
+    return frame_id
+
+
+def test_promote_defaults_to_reviewed_frames(client):
+    set_id, sub_class_id = _make_set_with_subclass(client)
+    reviewed_frame = _make_reviewed_frame_with_label(client, set_id, sub_class_id, frame_index=0)
+    _make_frame(client, frame_index=1)  # stays unlabeled, should not be promoted
+
+    resp = client.post(f"/sets/{set_id}/dataset-versions", json={"name": "v1"})
+    assert resp.status_code == 201
+    assert resp.json()["frame_count"] == 1
+
+    detail = client.get(f"/dataset-versions/{resp.json()['id']}").json()
+    assert [f["frame"]["id"] for f in detail["frames"]] == [reviewed_frame]
+    assert detail["frames"][0]["labels"][0]["sub_class_id"] == sub_class_id
+
+
+def test_promote_rejects_duplicate_name(client):
+    set_id, sub_class_id = _make_set_with_subclass(client)
+    _make_reviewed_frame_with_label(client, set_id, sub_class_id, frame_index=0)
+
+    client.post(f"/sets/{set_id}/dataset-versions", json={"name": "v1"})
+    resp = client.post(f"/sets/{set_id}/dataset-versions", json={"name": "v1"})
+    assert resp.status_code == 409
+
+
+def test_promote_rejects_empty_selection(client):
+    set_id, _ = _make_set_with_subclass(client)
+    resp = client.post(f"/sets/{set_id}/dataset-versions", json={"name": "v1"})
+    assert resp.status_code == 422
+
+
+def test_promote_with_explicit_frame_ids(client):
+    set_id, sub_class_id = _make_set_with_subclass(client)
+    frame_id = _make_frame(client, frame_index=0)  # not reviewed, but explicitly selected
+    client.post(
+        f"/frames/{frame_id}/sets/{set_id}/labels",
+        json={"sub_class_id": sub_class_id, "x": 0, "y": 0, "width": 1, "height": 1},
+    )
+
+    resp = client.post(f"/sets/{set_id}/dataset-versions", json={"name": "v1", "frame_ids": [frame_id]})
+    assert resp.status_code == 201
+    assert resp.json()["frame_count"] == 1
+
+
+def test_promote_rejects_unknown_frame_ids(client):
+    set_id, _ = _make_set_with_subclass(client)
+    resp = client.post(f"/sets/{set_id}/dataset-versions", json={"name": "v1", "frame_ids": [999]})
+    assert resp.status_code == 422
+
+
+def test_promote_requires_existing_set(client):
+    resp = client.post("/sets/999/dataset-versions", json={"name": "v1"})
+    assert resp.status_code == 404
+
+
+def test_dataset_version_is_frozen_after_label_edits(client):
+    set_id, sub_class_id = _make_set_with_subclass(client)
+    frame_id = _make_reviewed_frame_with_label(client, set_id, sub_class_id, frame_index=0)
+
+    version_id = client.post(f"/sets/{set_id}/dataset-versions", json={"name": "v1"}).json()["id"]
+    original_labels = client.get(f"/dataset-versions/{version_id}").json()["frames"][0]["labels"]
+    assert original_labels[0]["x"] == 1
+
+    # edit the live label after promotion
+    label_id = client.get(f"/frames/{frame_id}/sets/{set_id}/labels").json()[0]["id"]
+    client.patch(f"/labels/{label_id}", json={"x": 999})
+
+    # the promoted snapshot must be unaffected
+    frozen_labels = client.get(f"/dataset-versions/{version_id}").json()["frames"][0]["labels"]
+    assert frozen_labels[0]["x"] == 1
+
+    # deleting the live label must not touch the snapshot either
+    client.delete(f"/labels/{label_id}")
+    frozen_labels_after_delete = client.get(f"/dataset-versions/{version_id}").json()["frames"][0]["labels"]
+    assert len(frozen_labels_after_delete) == 1
+
+
+def test_list_dataset_versions(client):
+    set_id, sub_class_id = _make_set_with_subclass(client)
+    _make_reviewed_frame_with_label(client, set_id, sub_class_id, frame_index=0)
+    client.post(f"/sets/{set_id}/dataset-versions", json={"name": "v1"})
+
+    resp = client.get(f"/sets/{set_id}/dataset-versions")
+    assert resp.status_code == 200
+    assert [v["name"] for v in resp.json()] == ["v1"]
+    assert resp.json()[0]["frame_count"] == 1
+
+
+def test_get_missing_dataset_version(client):
+    resp = client.get("/dataset-versions/999")
+    assert resp.status_code == 404