32B-Para, n_perm 1; calmer Chinese example clips

app.py

@@ -10,7 +10,7 @@ from duplexjev import Question, default_table
 from duplexjev.table import LABELS
 
 MODEL_ID = os.environ.get("DUPLEXJEV_MODEL", "adventists-ai/DuplexJev-32B-Para")
-N_PERM = int(os.environ.get("DUPLEXJEV_NPERM", "2"))  # average over option orders: steadier answers
+N_PERM = int(os.environ.get("DUPLEXJEV_NPERM", "1"))  # 2 = average over option orders (same answers in our tests, 1.5x slower)
 GPU_SIZE = os.environ.get("DUPLEXJEV_GPU", "xlarge")  # 32B in bf16 needs the full 96 GB card
 MOCK = os.environ.get("DUPLEXJEV_MOCK") == "1"  # local UI test without a GPU
 
@@ -125,7 +125,7 @@ finished? should I stop talking? are they angry?* Every answer comes from **one
 """
 
 FOOT = """
-<sub>Model: DuplexJev-32B-Para (weights CC BY-NC 4.0); answers averaged over option orders. Audio is processed for this request only and not stored.
+<sub>Model: DuplexJev-32B-Para (weights CC BY-NC 4.0), one full ZeroGPU card. Audio is processed for this request only and not stored.
 Edit the questions freely: one per line, `name | question | option 1; option 2; …`.
 GPU time shown excludes the queue and network. 音频只用于本次判断，不保存。</sub>
 """

examples/zh_complete.wav

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4ff27bdeb094070c77d17ad56c3f8468fba8ad1dda046857e52cf8e97865c808
-size 313010
+oid sha256:a65087a2ff91209eccdf4094f41c318aa239618a047eaa8ffd9f243ae5f58024
+size 59772

examples/zh_happy.wav

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a9d483c777edb813972fb94642e0f5378d8110052b4badb4e2bb86ad7e3cef6a
-size 220850
+oid sha256:38b2590b82a4434d99f6c016aefd2a91fadb54419e3812367eb0124993fddd1a
+size 51924

examples/zh_wait.wav

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:98306d69c312f214898eba134ec93efb24d86c7c04a8f5fb42592b7829a11551
-size 190130
+oid sha256:cfcdd84db36bc6cea020fa78da8b2079fcd2fe8184187898e67ea7a23e7330ff
+size 123604