diff --git a/internal/uploads/intake.go b/internal/uploads/intake.go
index d7afd85..8f5ce13 100644
--- a/internal/uploads/intake.go
+++ b/internal/uploads/intake.go
@@ -58,6 +58,27 @@ func NewIntake(s Store, p Prober, sub Submitter, l Limits, tmpDir string) *Intak
 	return &Intake{store: s, prober: p, sub: sub, limits: l, tmpDir: tmpDir}
 }
 
+// ErrClientRead marks an error that came from READING the client's body: a connection that
+// dropped, a body cut short. It is the sender's to retry, unlike a spool, probe or database
+// fault, which is ours. It is a marker on the error, not a sentinel that others may wrap:
+// io.ErrUnexpectedEOF alone is also what a database drop looks like.
+var ErrClientRead = errors.New("uploads: client body read failed")
+
+// clientReader remembers the last non-EOF error its source returned, so Accept can tell
+// a failure reading the client from a failure writing the spool file.
+type clientReader struct {
+	r   io.Reader
+	err error
+}
+
+func (c *clientReader) Read(p []byte) (int, error) {
+	n, err := c.r.Read(p)
+	if err != nil && !errors.Is(err, io.EOF) {
+		c.err = err
+	}
+	return n, err
+}
+
 // busyMessage is what a user is told when the queue is full.
 const busyMessage = "Too many recordings are waiting to be transcribed. Please try again in a few minutes."
 
@@ -86,11 +107,15 @@ func (in *Intake) Accept(ctx context.Context, userID, title string, src io.Reade
 	}()
 
 	// Read one byte past the cap so "exactly at the limit" and "over" are told apart.
-	n, err := io.Copy(f, io.LimitReader(src, in.limits.MaxBytes+1))
+	cr := &clientReader{r: io.LimitReader(src, in.limits.MaxBytes+1)}
+	n, err := io.Copy(f, cr)
 	if cerr := f.Close(); err == nil {
 		err = cerr
 	}
 	if err != nil {
+		if cr.err != nil && errors.Is(err, cr.err) {
+			return "", fmt.Errorf("%w: %w", ErrClientRead, err)
+		}
 		return "", fmt.Errorf("uploads: receive: %w", err)
 	}
 	if n > in.limits.MaxBytes {