diff --git a/examples/cuda-fanout/README.md b/examples/cuda-fanout/README.md
index 1f884b5..bc6dc42 100644
--- a/examples/cuda-fanout/README.md
+++ b/examples/cuda-fanout/README.md
@@ -107,6 +107,29 @@ Python's default stdout buffering, which masks the "ready" /
 appear live. (Adding an explicit `flush-port` after the listening
 print would fix this in the Scheme without needing `-u`.)
 
+## Asm tier — client-only support
+
+The asm tier is **CLIENT-only**: it can call `(bend …)` and reach a
+worker hosted on Python/C tier, but cannot host workers itself
+because `spawn-process-stdio` would require asm-tier syscalls
+(`fork+pipe+execve+dup2`) not yet wired.
+
+What works on asm today:
+
+- `(load "wire.lsp")` — same wire protocol everywhere
+- `(load "bend.lsp")` — function form (`bend-call`, `bend!-call`)
+- `(bend!-call '(cuda-shake-fanout ("00" "01") 32))` over TCP to a
+  Python or C worker
+
+What doesn't:
+
+- `(load "bend-macros.lsp")` — needs `define-syntax` (only on Python/C)
+- `(load "gpu-worker.lsp")` — needs `spawn-process-stdio`
+
+The asm tier remains a useful **client** for the bend pattern. Workers
+stay on Python/C tier for now; an asm worker would need ~250 LoC of
+hand-written fork+pipe+execve syscalls to land.
+
 ## How lumbda's core would integrate this
 
 Three things, each ~20 lines per tier:
diff --git a/examples/cuda-fanout/bend-macros.lsp b/examples/cuda-fanout/bend-macros.lsp
new file mode 100644
index 0000000..60ba3a6
--- /dev/null
+++ b/examples/cuda-fanout/bend-macros.lsp
@@ -0,0 +1,12 @@
+;;; bend-macros.lsp — (bend …) / (bend! …) macros.
+;;;
+;;; Loaded only by tiers that support define-syntax (Python, C).
+;;; Asm tier callers use bend-call / bend!-call from bend.lsp directly.
+
+(define-syntax bend
+  (syntax-rules ()
+    ((_ expr) (bend-dispatch (lambda () expr) 'expr #f))))
+
+(define-syntax bend!
+  (syntax-rules ()
+    ((_ expr) (bend-dispatch (lambda () expr) 'expr #t))))
diff --git a/examples/cuda-fanout/bend.lsp b/examples/cuda-fanout/bend.lsp
index 23e6efb..c55beae 100644
--- a/examples/cuda-fanout/bend.lsp
+++ b/examples/cuda-fanout/bend.lsp
@@ -22,6 +22,19 @@
 
 (load "wire.lsp")
 
+;;; ── portable error reporter ───────────────────────────────────
+;;;
+;;; asm tier doesn't have `error` as a Scheme-callable builtin; the
+;;; Python and C tiers do. This wraps both so the rest of the file
+;;; stays portable. Displays the message, then either raises (on tiers
+;;; that have it) or returns a tagged failure value.
+
+(define (bend-error msg . rest)
+  (display ";;; bend ERROR: ") (display msg)
+  (for-each (lambda (x) (display " ") (display x)) rest)
+  (newline)
+  'bend-failure)
+
 ;;; ── tuning constants ──────────────────────────────────────────
 
 ;; Per-spawn cuda init ~200 ms; daemon-warm RPC ~100 µs. Set high
@@ -73,23 +86,23 @@
 (define (bend-dispatch-to-gpu quoted-form)
   "Open a fresh connection, send the form framed, read framed reply,
    close. Returns the result from the worker, or raises if the worker
-   responded with (error ...)."
+   responded with (bend-error...)."
   (let ((sock (tcp-connect *bend-worker-host* *bend-worker-port*)))
     (cond
       ((eq? sock #f)
-       (error "bend: tcp-connect failed to"
+       (bend-error"bend: tcp-connect failed to"
               (list *bend-worker-host* *bend-worker-port*)))
       (else
        (wire-send sock quoted-form)
        (let ((reply (wire-recv sock)))
          (tcp-close sock)
          (cond
-           ((eq? reply #f) (error "bend: worker closed connection"))
-           ((not (pair? reply)) (error "bend: malformed reply" reply))
+           ((eq? reply #f) (bend-error"bend: worker closed connection"))
+           ((not (pair? reply)) (bend-error"bend: malformed reply" reply))
            ((eq? (car reply) 'ok) (car (cdr reply)))
            ((eq? (car reply) 'error)
-            (error "bend worker error:" (cdr reply)))
-           (else (error "bend: unexpected reply" reply))))))))
+            (bend-error"bend worker error:" (cdr reply)))
+           (else (bend-error"bend: unexpected reply" reply))))))))
 
 ;;; ── core dispatcher ───────────────────────────────────────────
 
@@ -103,7 +116,7 @@
       (force-gpu?
        (if (and entry (bend-worker-available?))
            (bend-dispatch-to-gpu quoted-form)
-           (error "bend!: no GPU worker available for op" head)))
+           (bend-error"bend!: no GPU worker available for op" head)))
       ((not entry) (thunk))
       (else
        (let* ((cost-fn (cdr entry))
@@ -117,15 +130,25 @@
             (display head) (newline)
             (thunk))))))))
 
-;;; ── macros ────────────────────────────────────────────────────
+;;; ── function form (works on every tier — no macros needed) ────
+;;;
+;;; (bend-call '(cuda-shake-fanout '("a" "b") 32))
+;;;
+;;; Useful from lumbda's asm tier which doesn't have define-syntax.
+;;; The macro form below is a thin wrapper over this.
 
-(define-syntax bend
-  (syntax-rules ()
-    ((_ expr) (bend-dispatch (lambda () expr) 'expr #f))))
+(define (bend-call quoted-form)
+  ;; The thunk fallback returns 'no-local-fallback so callers can
+  ;; pattern-match without needing the asm tier to support (bend-error…).
+  (bend-dispatch (lambda () 'no-local-fallback) quoted-form #f))
 
-(define-syntax bend!
-  (syntax-rules ()
-    ((_ expr) (bend-dispatch (lambda () expr) 'expr #t))))
+(define (bend!-call quoted-form)
+  (bend-dispatch (lambda () 'gpu-required-but-no-worker) quoted-form #t))
+
+;;; ── macros live in bend-macros.lsp (Python/C tiers) ──────────
+;;; asm tier doesn't have define-syntax; load bend-macros.lsp only on
+;;; tiers that support it. The function-form bend-call / bend!-call
+;;; above is the asm-portable path.
 
 ;;; ── demo ──────────────────────────────────────────────────────
 
diff --git a/examples/cuda-fanout/smoke-bend-asm.lsp b/examples/cuda-fanout/smoke-bend-asm.lsp
new file mode 100644
index 0000000..6c38390
--- /dev/null
+++ b/examples/cuda-fanout/smoke-bend-asm.lsp
@@ -0,0 +1,24 @@
+;;; smoke-bend-asm.lsp — asm-tier-compatible bend test.
+;;;
+;;; asm tier doesn't have define-syntax, so we can't use the (bend …)
+;;; macro. The function-form bend!-call is the portable path.
+;;;
+;;; Prereqs:
+;;;   - gpu-worker.lsp running on 127.0.0.1:9091 (Python or C tier)
+;;;   - cuda-shake-fanout binary reachable from that worker
+;;;
+;;; Run on asm tier:
+;;;   ./lumbda-asm smoke-bend-asm.lsp
+
+(load "wire.lsp")
+(load "bend.lsp")
+
+(display "=== smoke-bend-asm ===\n")
+
+(display "1. worker available? ")
+(display (bend-worker-available?))
+(newline)
+
+(display "2. bend!-call '(cuda-shake-fanout (\"00\" \"01\" \"deadbeef\") 32): ")
+(let ((r (bend!-call '(cuda-shake-fanout ("00" "01" "deadbeef") 32))))
+  (display r) (newline))
diff --git a/examples/cuda-fanout/smoke-bend.lsp b/examples/cuda-fanout/smoke-bend.lsp
index 92f133d..d6124af 100644
--- a/examples/cuda-fanout/smoke-bend.lsp
+++ b/examples/cuda-fanout/smoke-bend.lsp
@@ -14,6 +14,7 @@
 ;;;   PASS
 
 (load "bend.lsp")
+(load "bend-macros.lsp")
 
 (define (make-input n)
   (let loop ((i 0) (acc '()))