diff --git a/src/picture.c b/src/picture.c
index f72828b..9739f6f 100644
--- a/src/picture.c
+++ b/src/picture.c
@@ -488,6 +488,49 @@ VAStatus RequestEndPicture(VADriverContextP context, VAContextID context_id)
 	if (rc != VA_STATUS_SUCCESS)
 		return rc;
 
+	/*
+	 * iter14 α-16: env-gated dump of OUTPUT bitstream bytes immediately
+	 * before QBUF. LIBVA_V4L2_DUMP_OUTPUT=<dir> writes source_data[0..
+	 * slices_size] to <dir>/output_<profile>_<surface>_<frame>.bin.
+	 * Discriminates whether libva writes the same H.264/HEVC slice bytes
+	 * as kdirect — if YES, Bug 4/5 are not in the OUTPUT-side; if NO,
+	 * narrow to which slice-write path produces the divergence.
+	 *
+	 * Off by default; no behavior change when env unset.
+	 */
+	{
+		static const char *dump_env = NULL;
+		static bool dump_env_checked = false;
+		if (!dump_env_checked) {
+			dump_env = getenv("LIBVA_V4L2_DUMP_OUTPUT");
+			dump_env_checked = true;
+		}
+		if (dump_env != NULL && dump_env[0] != '\0' &&
+		    surface_object->source_data != NULL &&
+		    surface_object->slices_size > 0) {
+			char path[256];
+			snprintf(path, sizeof(path),
+				 "%s/output_p%d_s%u_t%llu.bin",
+				 dump_env, (int)config_object->profile,
+				 (unsigned int)surface_object->base.id,
+				 (unsigned long long)context_object->timestamp_counter);
+			FILE *fp = fopen(path, "wb");
+			if (fp != NULL) {
+				size_t w = fwrite(surface_object->source_data,
+						  1, surface_object->slices_size,
+						  fp);
+				request_log("α-16: dumped %zu bytes to %s "
+					    "(slices_count=%u)\n",
+					    w, path,
+					    surface_object->slices_count);
+				fclose(fp);
+			} else {
+				request_log("α-16: fopen(%s) failed: %s\n",
+					    path, strerror(errno));
+			}
+		}
+	}
+
 	rc = v4l2_queue_buffer(driver_data->video_fd, -1, capture_type, NULL,
 			       surface_object->destination_index, 0,
 			       surface_object->destination_buffers_count);