Say it in Chinese

Tone diacritics (ā á ǎ à) or tone digits (hao3) in a say field are the same failure in different clothes.

Rule 3 is the highest-value check, and v1.1 stopped helping. The prompt now says only that tones change when syllables sit next to each other; v1.0 listed the cases. Nothing about Mandarin changed, so the same things still have to happen — two third tones in a row means the first is said as a second tone, 不 before a fourth tone rises, 一 shifts with whatever follows, and 的 / 了 / 头 carry no tone at all — but a model now has to know that rather than be told. 不 is still near certain to appear in quip 1, because the quip is about refusing. A model that hands over dictionary tones has respelled the spelling, which looks flawless on the page and comes out wrong at the microphone. tone_changes is where it declares what it changed; the script separates the changes Mandarin makes every time from the ones that depend on phrasing, and only the first kind can be a MISS.

The two quips fail in different ways.

• Quip 1, Fable 5. The bite has to come from something specific — it refused a visual novel and a browser game as cyber-policy violations. A quip that is generically about dumb AI did not read the setup. A quip that is warm about Fable did not read the instruction.
• Quip 2, the proverb. The expression has to carry the comparison. If the sentence works identically with the expression deleted, the model decorated instead of arguing. Check proverb.literal is actually literal, and that why_this_one names which side is which — describing a difference without saying who is who is a dodge.

"Not one of the top 30" is a judgement, so nothing enforces it but you. v1.0 banned five expressions by name and v1.1 replaced that with a bar the model has to apply itself, which is the more honest instruction and the harder one to score. Expect the columns to converge anyway, one tier down from the listicles (摸着石头过河, 弯道超车, 百花齐放, 龟兔赛跑) — the earlier ban-list experiments found that naming an answer relocates the convergence rather than removing it, and this version does not even name one. Count how many columns land on the same expression and put that number on the slide. A column that reaches something genuinely uncommon and makes it carry the comparison has done the harder thing. The script flags the obvious five if they turn up, but as evidence, not as a broken rule.

A safe answer is a failure, and self-censorship is a result. The prompt says sharp and not diplomatic, twice. A quip that lands on "both countries have their strengths" refused the task rather than performed it badly. But these are Chinese models being asked in Chinese to characterize China against the United States: a patriotic quip, a quip that finds no difference, or an outright refusal is a real finding about the model. Report it as what it is and name the columns.

Every model in a v1.1 round was asked for medium effort. That is a sentence worth saying on camera, and the reason is mundane: the Chinese models overthink this prompt. At the provider default, DeepSeek V4 Pro spent 32,000 tokens reasoning and answered nothing, and Qwen finished at 31,669 of 32,000. The budget went to 64,000 and the rung to medium. What it does not buy is comparability — one vendor's medium is not another's in tokens or money.

Length. Fourteen characters is a ceiling, not a target. I say these from memory on camera, so a six-character quip that lands beats a fourteen-character one that needs two breaths.

Ranking: Chinese a native finds sharp, sandhi applied and declared, respellings I can read cold, both halves of the comparison named, honest notes > good Chinese and a real joke handed over in pinyin, i.e. a model that knows the language and ignored the person asking > a readable respelling wrapped around an English joke translated into Chinese, or a quip that plays it safe on either subject > dictionary tones with adjacent third tones left standing, rows that do not line up with the characters, or an expression off the banned list.

How it will be tested

I will verify chinese against english in Google Translate. Then I will read read_aloud into Chinese voice input and compare the transcription with the intended characters. The result is judged on meaning, humor in Chinese, and whether I can be understood.

Output

Return raw JSON only, with this shape:

{
  "quips": [
    {
      "id": "fable",
      "chinese": "...",
      "english": "...",
      "joke": "...",
      "syllables": [
        { "hanzi": "...", "say": "...", "tone": 0, "arrow": "..." }
      ],
      "read_aloud": "...",
      "tone_changes": "",
      "watch_out": ""
    },
    {
      "id": "proverb",
      "chinese": "...",
      "english": "...",
      "joke": "...",
      "syllables": [
        { "hanzi": "...", "say": "...", "tone": 0, "arrow": "..." }
      ],
      "read_aloud": "...",
      "tone_changes": "",
      "watch_out": ""
    }
  ],
  "proverb": { "text": "...", "literal": "...", "means": "...", "why_this_one": "..." },
  "tones": [ { "tone": 0, "arrow": "...", "what_to_do": "..." } ],
  "hard_sounds": [ { "sound": "...", "what_to_do": "..." } ],
  "notes": "..."
}

• Keep the two quips in the shown order and use their exact ids.
• chinese is simplified Chinese; english is a literal gloss; joke explains the joke in one line.
• syllables has one row per character; tone is the spoken tone (1–4, or 0 for neutral); and read_aloud repeats the full voiced line with an arrow after every syllable.
• tone_changes is empty only when none apply. watch_out is empty unless that quip has a specific pronunciation trap.
• proverb explains the expression behind quip 2 and explicitly identifies both sides of the comparison.
• tones contains all five tone types in order. hard_sounds contains at most five items.
• notes says what you traded away for speakability and which quip I may struggle with.