Guard valuation questions asked in Chinese, Japanese or Korean

[zanwenfu]

Problem

"埃克森美孚值得买入还是持有？" ("is XOM worth buying or holding?") was not recognised as a valuation question, because the detection pattern was English only. The publication guard never ran.

The model's draft was published as written: "已有持仓可继续持有" ("keep holding if you own it"). There was no NOT RATED, for a company the engine declines to value. This predates #24.

Change

• Recognition. The chat agent now recognises valuation questions in Chinese, Japanese and Korean: buy/sell/hold decisions, valuation, price targets and analysis, the same things the English pattern looks for. These questions now get the fixed statement plus the analysis section from Keep the run's analysis when the guard answers with its fixed statement #24, whose two checks already work in any language.
• Published headlines. A broad answer written in Chinese, Japanese or Korean cannot be checked against a published headline, because the conflict patterns only read English. "报告评级为卖出" ("the report's rating is sell") passed while the report said HOLD. Such an answer now gets the headline itself plus the checked analysis.
• Chinese first. A question asked in Chinese now reads its position in Chinese first. This is one deterministic line built from what the guard recorded, for example "XOM：未评级。不发布评级或公允价值：其现金流随大宗商品价格波动……" ("XOM: not rated. No rating or fair value is published: its cash flows follow commodity prices…"). The English statement follows.
  • The guard now records which statement it used: refused, withheld or published.
  • Each refusal reason has a Chinese version in summary_evidence.
• Language of the analysis. The analysis turn now names the language ("Write it in Chinese."). When the instruction only said "the user's language", Chinese questions got English sections, and the language check refused them.
• Chinese claim terms. Chinese value and target terms (目标价, 内在价值, 上行空间 …) count as claims only with a figure in the same clause, like their English equivalents. A disclaimer such as "无法据此量化…内在价值" ("intrinsic value cannot be quantified from this") had caused a whole section to be refused.

Verification

• 1561 tests pass, and 5 mutations were each caught.
• Production vs candidate on real questions:
  • Chinese valuation questions (XOM refused, META withheld, MSFT published, TSLA analyze): all are now guarded. Each reads its position in Chinese first and carries a Chinese analysis that passed both checks. I read each one by hand.
  • Unchanged: the Chinese news question and the English questions behave as before.
• Valuation canary: PASS, with no values changed.

🤖 Generated with Claude Code