fix: harden general consultation output guard

This commit is contained in:
Jesse_Chen
2026-07-21 10:11:05 +08:00
parent 0edb36e26b
commit 1b3fdcd641
2 changed files with 103 additions and 18 deletions
+66 -18
View File
@@ -11,30 +11,78 @@ const guaranteeConclusionPatterns = [
/(?:^|[.?!\n])[^.?!\n]*\b(?:will definitely|guaranteed? to|certain to|without doubt)\b[^.?!\n]*/gi,
];
// The boundary is structural: first require a personalized subject, then a
// chart object and a placement/conclusion predicate. Planet names are a finite
// vocabulary supplement, not the primary detection mechanism.
const chineseChartObjectPattern = /(?:?|||||D\s*\d+|(?:)?|[\p{Script=Han}A-Za-z0-9]{0,8}||||||||||||Rahu|Ketu)/iu;
const chineseChartConclusionPattern = /(?:(?:||)?||||||||||?|?|||||[:])/iu;
const englishChartObjectPattern = /\b(?:natal\s+chart|birth\s+chart|chart|ascendant|rising\s+sign|D\s*\d+|(?:\d+(?:st|nd|rd|th)|[a-z]+)\s+house|house|planet|sun|moon|mars|mercury|jupiter|venus|saturn|rahu|ketu)\b/iu;
const englishChartConclusionPattern = /(?:\b(?:is|are|falls?|lands?|sits?|placed?|located?|shows?|indicates?|means?|rules?|strong|weak)\b|[:])/iu;
const chinesePossessiveChartSubjectPattern = /(?:|)\s*(?:|(?:)?)\s*(?:?|||||D\s*\d+|(?:)?|[\p{Script=Han}A-Za-z0-9]{0,8}||||||||||||Rahu|Ketu)/iu;
const chineseBareChartSubjectPattern = /(?:|)\s*(?:?|||||D\s*\d+|(?:)?|(?:\s*)?[0-9]+\s*||||||||||||Rahu|Ketu)/iu;
const chineseChartAddressesUserPattern = /(?:?|||||D\s*\d+|(?:)?)\s*(?:|||)\s*(?:|)/iu;
const englishPersonalChartSubjectPattern = /\b(?:your|the\s+user(?:'s)?)\s+(?:personal\s+)?(?:natal\s+chart|birth\s+chart|chart|ascendant|rising\s+sign|D\s*\d+|(?:\d+(?:st|nd|rd|th)|[a-z]+)\s+house|house|planet|sun|moon|mars|mercury|jupiter|venus|saturn|rahu|ketu)\b/iu;
const englishChartAddressesUserPattern = /\b(?:the\s+)?(?:natal\s+|birth\s+)?chart\s+(?:shows?|indicates?|means?)\s+(?:that\s+)?you\b/iu;
// The boundary is structural: an educational question frame is removed first,
// then a personal subject/context, chart object, and chart predicate must agree.
// This avoids treating every sentence that merely contains `你` and `宫` as a
// personal chart claim.
const chineseZodiacSignSource = String.raw`(?:白羊|金牛|双子|巨蟹|狮子|处女|天秤|天蝎|射手|摩羯|水瓶|双鱼)`;
const chineseNamedHouseSource = String.raw`(?:命|财帛|兄弟|田宅|子女|夫妻|婚姻|疾厄|迁移|事业|官禄|交友|仆役|福德|父母)宫`;
const chineseChartObjectSource = String.raw`(?:盘面?|星盘|命盘|出生盘|本命盘|D\s*\d+|上升(?:星座)?|星座|${chineseZodiacSignSource}(?:座|上升)|(?:第\s*)?[一二三四五六七八九十百0-9]+\s*宫|${chineseNamedHouseSource}|宫位|行星|星体|太阳|月亮|火星|水星|木星|金星|土星|罗喉|凯图|Rahu|Ketu)`;
const chineseChartObjectPattern = new RegExp(chineseChartObjectSource, "iu");
const chineseChartPredicatePattern = /(?:(?:||)?|||||||||||?|?|||||[:])/iu;
const chineseOwnedChartSubjectPattern = new RegExp(
String.raw`(?:你|您)\s*(?:的|个人(?:的)?)?\s*${chineseChartObjectSource}`,
"iu",
);
const chineseUserHasChartPattern = new RegExp(
String.raw`(?:你|您)\s*(?:为|是|有)\s*(?:一(?:个|颗)\s*)?${chineseChartObjectSource}`,
"iu",
);
const chinesePersonalChartContextPattern = /(?:(?:|)|(?:|)(?:)?(?:?||||)|(?:|)(?:)?(?:?||||)(?:|)?)/iu;
const chineseChartAddressesUserPattern = new RegExp(
String.raw`${chineseChartObjectSource}\s*(?:显示|表明|说明|意味着)[^。!?.!?\n]{0,24}(?:你|您)`,
"iu",
);
const englishZodiacSignSource = String.raw`(?:aries|taurus|gemini|cancer|leo|virgo|libra|scorpio|sagittarius|capricorn|aquarius|pisces)`;
const englishChartObjectSource = String.raw`(?:natal\s+chart|birth\s+chart|chart|ascendant|rising(?:\s+sign)?|D\s*\d+(?:\s+chart)?|placement|(?:\d+(?:st|nd|rd|th)|[a-z]+)\s+house|house|planet|sun|moon|mars|mercury|jupiter|venus|saturn|rahu|ketu)`;
const englishChartObjectPattern = new RegExp(String.raw`\b${englishChartObjectSource}\b`, "iu");
const englishChartPredicatePattern = /(?:\b(?:has|have|is|are|occup(?:y|ies)|falls?|lands?|sits?|placed?|located?|shows?|indicates?|means?|rules?|strong|weak)\b|[:])/iu;
const englishOwnedChartSubjectPattern = new RegExp(
String.raw`\b(?:your|the\s+user(?:'s)?)\s+(?:personal\s+)?${englishChartObjectSource}\b`,
"iu",
);
const englishUserHasChartPattern = new RegExp(
String.raw`\byou\s+have\s+(?:an?\s+)?(?:${englishZodiacSignSource}\s+rising|${englishChartObjectSource})\b`,
"iu",
);
const englishPersonalChartContextPattern = /\b(?:for\s+you|(?:in|from|based\s+on|according\s+to)\s+your\s+(?:natal\s+|birth\s+)?chart)\b/iu;
const englishChartAddressesUserPattern = new RegExp(
String.raw`\b${englishChartObjectSource}\s+(?:shows?|indicates?|means?)\s+(?:that\s+)?you\b`,
"iu",
);
const chineseEducationalFramePattern = /^(?:(?:|)(?:)?(?:|)|(?:|)(?:)?(?:)?(?:)?)[\s,:]*/iu;
const chineseAboutGeneralMeaningPattern = /^\s*(?:|)(?:)?\s*(.+?(?:|))$/iu;
const englishEducationalFramePattern = /^(?:your\s+question\s+is|you\s+asked\s+about)\s*/iu;
function normalizeClaimClause(clause: string) {
let normalized = clause
.normalize("NFKC")
.replace(/[*_`~]+/gu, "")
.replace(/\s+/gu, " ")
.trim();
normalized = normalized
.replace(chineseEducationalFramePattern, "")
.replace(englishEducationalFramePattern, "");
const aboutGeneralMeaning = chineseAboutGeneralMeaningPattern.exec(normalized);
return aboutGeneralMeaning?.[1] ?? normalized;
}
function isPersonalChartConclusion(clause: string) {
const normalized = clause.normalize("NFKC");
const chineseStructure = (chinesePossessiveChartSubjectPattern.test(normalized)
|| chineseBareChartSubjectPattern.test(normalized)
const normalized = normalizeClaimClause(clause);
const chineseStructure = (chineseOwnedChartSubjectPattern.test(normalized)
|| chineseUserHasChartPattern.test(normalized)
|| chinesePersonalChartContextPattern.test(normalized)
|| chineseChartAddressesUserPattern.test(normalized))
&& chineseChartObjectPattern.test(normalized)
&& chineseChartConclusionPattern.test(normalized);
const englishStructure = (englishPersonalChartSubjectPattern.test(normalized)
&& chineseChartPredicatePattern.test(normalized);
const englishStructure = (englishOwnedChartSubjectPattern.test(normalized)
|| englishUserHasChartPattern.test(normalized)
|| englishPersonalChartContextPattern.test(normalized)
|| englishChartAddressesUserPattern.test(normalized))
&& englishChartObjectPattern.test(normalized)
&& englishChartConclusionPattern.test(normalized);
&& englishChartPredicatePattern.test(normalized);
return chineseStructure || englishStructure;
}
@@ -109,6 +109,43 @@ test("general mode structurally rejects personalized chart placements in Chinese
);
});
test("general mode rejects personalized chart predicates beyond possessive placements", () => {
const unsafeClaims = [
"您的上升星座为巨蟹。",
"你是巨蟹上升。",
"您有一个巨蟹上升。",
"You have Cancer rising.",
"In your chart, Venus occupies the seventh house.",
"For you, Venus falls in the seventh house.",
];
for (const claim of unsafeClaims) {
const guarded = guardGeneralNoBirthTimeOutput(claim);
const punctuation = claim.endsWith("。") ? "。" : ".";
assert.equal(guarded, `${GENERAL_NO_BIRTH_TIME_REFUSAL}${punctuation}`, claim);
}
});
test("general mode preserves educational chart questions and general knowledge", () => {
const safeStatements = [
"你的问题是第七宫的一般含义。",
"你问的是第七宫的一般含义。",
"你问的第七宫在占星概念中常与关系相关。",
"关于你的第七宫的一般含义。",
"你是一个正在研究第七宫的学生。",
"你可以先了解第七宫的一般概念。",
"第七宫在占星概念中常与关系相关。",
"Your question is the general meaning of the seventh house.",
"You asked about the seventh house.",
"You have a question about the seventh house.",
"The seventh house is associated with relationships.",
];
for (const statement of safeStatements) {
assert.equal(guardGeneralNoBirthTimeOutput(statement), statement);
}
});
test("hidden AYANAM comments cannot split a personalized claim around the guard", async () => {
const title = "<!--AYANAM_TITLE:一般占星咨询-->";
const suggestions = '<!--AYANAM_SUGGESTIONS:["了解第七宫的一般概念","先完成生时校正","改问一般知识"]-->';