diff --git a/internal/engine/engine.go b/internal/engine/engine.go
index 1968c82..b7b61fc 100644
--- a/internal/engine/engine.go
+++ b/internal/engine/engine.go
@@ -2,6 +2,7 @@ package engine
 
 import (
 	"context"
+	"encoding/json"
 	"fmt"
 	"io"
 	"strconv"
@@ -9,6 +10,7 @@ import (
 
 	"github.com/docker/docker/api/types/container"
 	"github.com/docker/docker/api/types/filters"
+	"github.com/docker/docker/api/types/image"
 	"github.com/docker/docker/client"
 	"github.com/docker/docker/pkg/stdcopy"
 	"github.com/docker/go-connections/nat"
@@ -98,6 +100,9 @@ func toDockerConfig(s Spec, bindHost, gpuDriver string) (*container.Config, *con
 }
 
 func (d *Docker) Start(ctx context.Context, s Spec) (string, error) {
+	if err := d.ensureImage(ctx, s.Image); err != nil {
+		return "", err
+	}
 	cfg, host := toDockerConfig(s, d.bindHost, d.gpuDriver)
 	created, err := d.cli.ContainerCreate(ctx, cfg, host, nil, nil, s.Name)
 	if err != nil {
@@ -109,6 +114,66 @@ func (d *Docker) Start(ctx context.Context, s Spec) (string, error) {
 	return created.ID, nil
 }
 
+// ensureImage pulls ref if it is not already present locally. The Engine
+// API's ContainerCreate, unlike `docker run`, never pulls a missing image
+// itself - it fails outright with "No such image". This went unnoticed
+// through every deploy this project has ever made, because every node
+// Sous had run on already had its images cached (from the single-node
+// Sous era, or from this fleet's own precedent of pre-pulling vLLM images
+// by hand). It surfaced for real on aorus-ubuntu's first-ever deployment:
+// a genuinely cold Docker install, with nothing cached at all.
+//
+// Checks local presence FIRST rather than pulling unconditionally on
+// every call - matching `docker run`'s own default ("pull if missing",
+// not "pull always") and avoiding a needless registry round-trip on the
+// overwhelmingly common case where the image core recipes (which pin
+// digests, not floating tags) already have cached.
+func (d *Docker) ensureImage(ctx context.Context, ref string) error {
+	if _, err := d.cli.ImageInspect(ctx, ref); err == nil {
+		return nil
+	} else if !client.IsErrNotFound(err) {
+		return fmt.Errorf("engine: inspect image %s: %w", ref, err)
+	}
+
+	rc, err := d.cli.ImagePull(ctx, ref, image.PullOptions{})
+	if err != nil {
+		return fmt.Errorf("engine: pull image %s: %w", ref, err)
+	}
+	defer rc.Close()
+
+	if err := drainPullStream(rc); err != nil {
+		return fmt.Errorf("engine: pull image %s: %w", ref, err)
+	}
+	return nil
+}
+
+// drainPullStream reads an ImagePull response to completion, which is
+// required for the pull to actually finish (it is not synchronous until
+// the stream is read) - and, unlike a plain io.Copy(io.Discard, r), also
+// catches a registry-side failure (auth, missing manifest, ...), which
+// Docker reports as an "error" field INSIDE this JSON stream rather than
+// as a Go error from ImagePull itself. Split out from ensureImage so the
+// decode/error-detection logic - the actual subtle part - is testable
+// against a crafted byte stream, with no real Docker daemon or network
+// access required.
+func drainPullStream(r io.Reader) error {
+	dec := json.NewDecoder(r)
+	for {
+		var msg struct {
+			Error string `json:"error"`
+		}
+		if err := dec.Decode(&msg); err != nil {
+			if err == io.EOF {
+				return nil
+			}
+			return fmt.Errorf("reading progress: %w", err)
+		}
+		if msg.Error != "" {
+			return fmt.Errorf("%s", msg.Error)
+		}
+	}
+}
+
 // Stop stops and removes. Leaving a stopped container behind would make the
 // next create fail on the name, which is the kind of failure that gets
 // misread as a problem with the new model.
diff --git a/internal/engine/engine_test.go b/internal/engine/engine_test.go
index f376fb4..838e79f 100644
--- a/internal/engine/engine_test.go
+++ b/internal/engine/engine_test.go
@@ -1,6 +1,8 @@
 package engine
 
 import (
+	"context"
+	"strings"
 	"testing"
 
 	"github.com/docker/go-connections/nat"
@@ -116,6 +118,80 @@ func TestNoEntrypointLeavesImageDefault(t *testing.T) {
 	}
 }
 
+// A real docker pull's progress stream, one JSON object per line, no
+// embedded error - captured in shape from an actual `docker pull` (status,
+// progressDetail, id fields), not invented.
+const realPullStreamNoError = `{"status":"Pulling from library/alpine","id":"3.21"}
+{"status":"Pulling fs layer","progressDetail":{},"id":"9b18e9b68314"}
+{"status":"Downloading","progressDetail":{"current":1024,"total":3072},"progress":"[====>   ]  1024B/3072B","id":"9b18e9b68314"}
+{"status":"Download complete","progressDetail":{},"id":"9b18e9b68314"}
+{"status":"Pull complete","progressDetail":{},"id":"9b18e9b68314"}
+{"status":"Digest: sha256:abc123"}
+{"status":"Status: Downloaded newer image for alpine:3.21"}
+`
+
+func TestDrainPullStreamSucceedsOnARealNoErrorStream(t *testing.T) {
+	if err := drainPullStream(strings.NewReader(realPullStreamNoError)); err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+}
+
+func TestDrainPullStreamCatchesAnErrorEmbeddedMidStream(t *testing.T) {
+	// The exact bug ensureImage exists to avoid: a registry-side failure
+	// (bad ref, auth, ...) arrives as an "error" field INSIDE the JSON
+	// stream, after several genuine progress lines - not as a Go error
+	// from ImagePull itself. A plain io.Copy(io.Discard, r) would drain
+	// this to EOF and report success.
+	stream := `{"status":"Pulling from library/alpine","id":"3.21"}
+{"status":"Pulling fs layer","progressDetail":{},"id":"9b18e9b68314"}
+{"errorDetail":{"message":"manifest unknown: manifest unknown"},"error":"manifest unknown: manifest unknown"}
+`
+	err := drainPullStream(strings.NewReader(stream))
+	if err == nil {
+		t.Fatal("expected an error, got nil")
+	}
+	if !strings.Contains(err.Error(), "manifest unknown") {
+		t.Fatalf("error should surface the registry's own message, got: %v", err)
+	}
+}
+
+func TestDrainPullStreamHandlesAnEmptyStream(t *testing.T) {
+	if err := drainPullStream(strings.NewReader("")); err != nil {
+		t.Fatalf("unexpected error on an empty stream: %v", err)
+	}
+}
+
+func TestDrainPullStreamSurfacesMalformedJSON(t *testing.T) {
+	err := drainPullStream(strings.NewReader("{not json"))
+	if err == nil {
+		t.Fatal("expected an error for malformed JSON, got nil")
+	}
+}
+
+// TestEnsureImageSkipsAnAlreadyCachedImage is a real integration test
+// against an actual Docker daemon, not a fake - ensureImage's whole
+// premise (check ImageInspect before ever calling ImagePull) isn't
+// meaningfully testable through drainPullStream alone, since that
+// covers only what happens once a pull stream exists. Skips cleanly if
+// no daemon is reachable or the fixture image isn't cached, rather than
+// failing a CI environment without Docker access - matching this
+// project's existing pattern of degrading gracefully for environment
+// limits (e.g. -race being unavailable in sandboxes without a C
+// toolchain) rather than papering over the gap with a fake.
+func TestEnsureImageSkipsAnAlreadyCachedImage(t *testing.T) {
+	d, err := New("", "cdi")
+	if err != nil {
+		t.Skipf("no local Docker daemon reachable: %v", err)
+	}
+	const fixtureImage = "alpine:3.21"
+	if _, err := d.cli.ImageInspect(context.Background(), fixtureImage); err != nil {
+		t.Skipf("fixture image %s not already cached locally: %v", fixtureImage, err)
+	}
+	if err := d.ensureImage(context.Background(), fixtureImage); err != nil {
+		t.Fatalf("ensureImage on an already-cached image should not error: %v", err)
+	}
+}
+
 func TestBindsAndRestartPolicy(t *testing.T) {
 	s := Spec{Name: "n", Image: "i", ContainerPort: 8000,
 		Binds: []string{"/models:/root/.cache/huggingface"}}
diff --git a/internal/engine/job.go b/internal/engine/job.go
index 22adf12..620f9e8 100644
--- a/internal/engine/job.go
+++ b/internal/engine/job.go
@@ -53,6 +53,9 @@ type JobSpec struct {
 // minutes; holding a request open for that is the same mistake undeploy used to
 // make. Progress is read afterwards from the container's own state and logs.
 func (d *Docker) StartJob(ctx context.Context, s JobSpec) (string, error) {
+	if err := d.ensureImage(ctx, s.Image); err != nil {
+		return "", err
+	}
 	cfg := &container.Config{
 		Image:  s.Image,
 		Cmd:    s.Cmd,