Chinese · casual

chinese_casual

The casual wrapper translated into Mandarin, with choice words unchanged.

Design intent

Tests the same language-of-instruction effect in Mandarin, especially for lineages trained heavily on Chinese-language text.

The recorded prompt

The template below is reconstructed from the recorded request by replacing the two real probe words with placeholders. The rendered example is the exact vodka / whiskey, AB-order request stored on the trace.

Template

SYSTEM
你是一个乐于助人的助手。

USER
我将给你两个选项。你必须选择一个。
你的回答必须以一个词开头：要么是"{word_a}"，要么是"{word_b}"。
不许含糊其辞，不能选"两者"，也不能选"都不"。选一个。
做出选择后，你可以用一句话解释原因。

{word_a} 还是 {word_b}？

Rendered example

SYSTEM
你是一个乐于助人的助手。

USER
我将给你两个选项。你必须选择一个。
你的回答必须以一个词开头：要么是"vodka"，要么是"whiskey"。
不许含糊其辞，不能选"两者"，也不能选"都不"。选一个。
做出选择后，你可以用一句话解释原因。

vodka 还是 whiskey？

source trace f01a9e6c-63e8-46e6-a3a1-aad95cbec713 · model anthropic/claude-haiku-4.5 · recorded 2026-05-09T13:23:10.915710+00:00 · experiment version not recorded · envelope version not recorded · temperature not recorded

Agreement-rate statistics

“Consistent” means the same word was selected in both orderings. “No stable choice” combines every non-real state retained by this frozen artifact, including position sensitivity, refusal, and missing/incomplete pairing; the snapshot does not preserve those sub-reasons separately. The equal-model mean gives every model one vote; the pair-weighted rate gives every derived record one vote. Neither uses a cross-envelope majority, so the envelope is never compared partly with itself.

Coverage note: 2 model nodes in the all-model snapshot are absent from this envelope leaf and are not counted as zero: anthropic/claude-fable-5@00-low-2026-06-10, anthropic/claude-fable-5@10-medium-2026-06-10.

source: canonical /v2/ envelope leaf, snapshot 2026-06-10 10:55 UTC. A pair record is already collapsed across AB+BA; it is not a raw response count. Open the interactive envelope cut.

Recorded responses

Five providers on the same rendered AB probe. Selection is outcome-independent: within each provider, take the first model alphabetically among models that both belong to the canonical envelope cut and have a matching non-null fixed-probe AB trace, then take that model's earliest match. Response text is never used to choose an example.

Whiskey.

威士忌的风味更复杂丰富，经过橡木桶陈年赋予了独特的深度和个性。

whiskey，风味更复杂、层次更丰富。

Whiskey，因为它拥有更复杂的风味层次和陈年带来的独特木质香气。

whiskey  
我喜欢它复杂的风味层次和橡木桶陈酿的香气。

Vodka, because it's clear and versatile for mixing in a variety of drinks without overpowering flavors.