From 41c1fc3ccd8d79e5ebefc24ed89832d650fe83d2 Mon Sep 17 00:00:00 2001 From: Claude Date: Sat, 21 Nov 2026 08:07:39 +0800 Subject: [PATCH] fix: normal readings never calibrate on the user's past events (BUG-2111) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The astrologer (2026-11-21) ruled that ordinary readings and forecasts never use the user's experiences or feedback; only the rectification flow may. Her reference files still told the model to back-calculate the user's known events first, stop predicting below 3/5 matches, and grade A by such verification. Until she edits them upstream, a product layer cuts those passages out of every text sent to a model: the personal reading's quoted references, the chat's methodology excerpts, and every chat prompt before it reaches the provider (covering skill_read). The files are not edited; a test names any rule that stops matching once she changes them. Grade A no longer asks for a historical back-test. The status user_history_verification_required is no longer a gap: no verification questions, no penalty in ranking, never shown as 「需要过往经历核对」, not offered to the report or synastry writers. 「未见阻断项」 is caught in the reading and in report-question answers. Contract §AU announces the layer. Co-Authored-By: Claude Opus 5.5 Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8 --- docs/BUG_HISTORY.md | 17 +++ frontend/src/lib/annual-question-check.ts | 2 + frontend/src/lib/consultation-methodology.ts | 4 +- .../src/lib/consultation-thinking-plan.ts | 2 +- .../src/lib/personal-report-generation.ts | 8 +- .../src/lib/personal-report-reading-terms.ts | 5 +- frontend/src/lib/personal-report-reading.ts | 2 +- frontend/src/lib/report-fact-table-display.ts | 5 +- frontend/src/lib/report-fact-tables.ts | 4 +- frontend/src/lib/synastry-full-generate.ts | 5 +- frontend/src/lib/synastry-full-sections.ts | 11 +- frontend/src/lib/synastry-full-timeline.ts | 5 +- frontend/src/lib/user-history-isolation.ts | 144 ++++++++++++++++++ frontend/src/mastra/personal-report.ts | 4 +- frontend/src/mastra/product-voice.ts | 4 +- frontend/src/mastra/report-reading-skill.ts | 8 +- frontend/src/mastra/skill-binding.ts | 16 +- .../tests/consult-chat-reopt-20261107.test.ts | 5 +- frontend/tests/personal-report-api.test.ts | 8 +- .../personal-report-generation-v2.test.ts | 9 +- frontend/tests/report-fact-tables.test.ts | 4 +- .../tests/synastry-exploratory-2014.test.tsx | 4 +- .../tests/user-history-isolation-2111.test.ts | 133 ++++++++++++++++ references/product_runtime_contract.md | 11 ++ 24 files changed, 389 insertions(+), 31 deletions(-) create mode 100644 frontend/src/lib/user-history-isolation.ts create mode 100644 frontend/tests/user-history-isolation-2111.test.ts diff --git a/docs/BUG_HISTORY.md b/docs/BUG_HISTORY.md index e602b2457..b67e34065 100644 --- a/docs/BUG_HISTORY.md +++ b/docs/BUG_HISTORY.md @@ -28473,3 +28473,20 @@ - 验证与防复发:`tests/test_report_leftovers_2109.py`(18 例)在两张公开盘的中英文报告上断言:英文版没有保留标记和中文行,Pada 行不印「未列」,相位矩阵没有罗睺—计都、没有原始阶段词、英文标题干净且覆盖对照表能找到,资料章节相位表里这一对不再标成相位,类别栏没有原始键,宫主表头是中文。`tests/test_report_fields_2065a.py` 里钉住旧写法的一处按原值、新值、原因更新。改前在干净检出上 17 例全部失败。 - 关联记录:BUG-2065(相位表)、BUG-2095 / BUG-2104(读者用词)、BUG-2105(Arudha 表)。 - 修复版本:`claude/report-leftovers-2109`。 + +## BUG-2111 | 普通解读和推运送给模型的材料里仍要求用用户过往经历校准 + +- 状态:fixed-pending-verify +- 现象:占星师 2026-11-21 审核发现,送给模型的 `references/precision-reading-methodology.md` 不只一处要求先用用户已知的过去事件倒推校准、匹配不足就停止预测,并把「A-已验证」定义为经用户已知事件倒推验证;同类说法还在 `ai-reading-workflow-prompt.md` §5.0.1–5.0.2(含「用户未提供过往事件就必须降低置信度」)和 `transit-actionable-output-guide.md` 要素3。产品自己的提示词里也有:等级 A「个人预测最好有历史回测」,状态规则要求 `user_history_verification_required`「写成待核验并给出可核验的问题」,个人报告卡片带「建议结合真实经历核验本主题的部分证据结论」,事实表显示「需要过往经历核对」,合盘写作可选这个标签并显示「需要用真实经历核对」。 +- 触发条件:任何普通解读(个人报告解读整段引用方法论文件;问答摘录行运规范、并可用 `skill_read` 打开任何参考文件)。 +- 根因:占星师的参考文件把用户经历校准写成通用流程;产品此前只在问答的反馈处理里隔离了用户反馈(BUG-1980),没有隔离送给模型的方法原文,也没有清理把等级和状态挂到用户经历上的产品文字。 +- 修复(模型输入一侧;产生这个状态的 Python 一侧见 BUG-2110): + - 新模块 `frontend/src/lib/user-history-isolation.ts`:按原文内容(不按行号)找出 10 处段落,整段换成一句产品说明、单行直接去掉;参考文件本身不改。用在个人报告解读的参考文件引用(`report-reading-skill.ts`,记录的哈希仍是原文件的)、问答方法摘录的文件读取(`consultation-methodology.ts`),以及问答每次发给模型前的整份提示(`skill-binding.ts` 的 `processLLMRequest`,覆盖 `skill_read` 读到的任何文件,不改存下来的工具结果)。生时纠正流程不读这些文件。 + - `product-voice.ts`:A 不再以历史回测为条件;等级只看盘面依据,不向用户要过往经历。`consultation-thinking-plan.ts`:`user_history_verification_required` 是旧数据标记、不算缺口,不写待核验、不提核验问题、不降结论。 + - 个人报告:事实表把这个状态按未声明处理;显示标签改成中性;状态排序与 `parameter_sensitive` 同级;不再生成「结合真实经历核验」的问题;写作代理的输出选项里去掉这个状态(存下的旧报告仍按 `CLAIM_STATUSES` 校验通过)。合盘:写作不再提供这个标签,显示词改中性,候选事件的说明不再要求用两人经历核对。 + - 「未见阻断项」(`no_open_blocker`):解读检查和报告内提问的用词检查会拦下,走现有的一次修复或改写。 + - `references/product_runtime_contract.md` §AU 记录这一产品隔离及其临时性。 +- 验证:`frontend/tests/user-history-isolation-2111.test.ts`(7 例:每条规则仍能匹配上游原文——她改完后会失败并点名;隔离后无校准句、其余方法原样;解读提示;问答处理器;行运规范摘录;等级和状态文字;未见阻断项);`synastry-exploratory-2014` 按原值 / 新值 / 原因更新。 +- 防复发:上游段落一变,匹配测试点名失败;送给模型的三条路径都走同一个模块。 +- 关联记录:BUG-1980(问答反馈隔离)、BUG-2110(Python 一侧的状态)、BUG-2098 / BUG-2101(同一轮审核)。 +- 修复版本:`claude/history-isolation-2111`。 diff --git a/frontend/src/lib/annual-question-check.ts b/frontend/src/lib/annual-question-check.ts index e38d4a7b3..b4c3bc39b 100644 --- a/frontend/src/lib/annual-question-check.ts +++ b/frontend/src/lib/annual-question-check.ts @@ -77,6 +77,8 @@ const ANNUAL_INTERNAL: ReadonlyArray = [ ["证据等级", /证据等级|把握等级|把握上限|置信度/u], ["等级字母", /(?:把握|等级|置信)[^,。;\n]{0,6}[A-D](?![A-Za-z])|[A-D]\s*[级档]/u], ["偏顺 / 偏紧", /偏顺|偏紧/u], + // BUG-2111 (astrologer 2026-11-21): 「未见阻断项」 is an internal state, not evidence sufficiency. + ["未见阻断项", /未见阻断项|no[ _]open[ _]blocker/iu], ["Mudda", /Mudda/iu], ["Tajika", /Tajika/iu], ["Varshaphala", /Varshaphala/iu], diff --git a/frontend/src/lib/consultation-methodology.ts b/frontend/src/lib/consultation-methodology.ts index c26988bd2..0ca7b82b9 100644 --- a/frontend/src/lib/consultation-methodology.ts +++ b/frontend/src/lib/consultation-methodology.ts @@ -1,4 +1,5 @@ import { existsSync, readFileSync, statSync } from "node:fs"; +import { isolateUserHistoryCalibration } from "./user-history-isolation.ts"; import { resolve } from "node:path"; import { ANNUAL_BACKEND_RULES, ANNUAL_BACKEND_RULES_SOURCE } from "./annual-backend-rules.ts"; import { z } from "zod"; @@ -557,7 +558,8 @@ function packageFile(relativePath: string): string | null { if (hit !== undefined) return hit; let content: string | null = null; try { - content = readFileSync(resolvePackageFile(relativePath), "utf8"); + // BUG-2111: passages that calibrate on the user's past events never reach the model. + content = isolateUserHistoryCalibration(readFileSync(resolvePackageFile(relativePath), "utf8")).text; } catch { content = null; } diff --git a/frontend/src/lib/consultation-thinking-plan.ts b/frontend/src/lib/consultation-thinking-plan.ts index 8dc6f0726..d752a1693 100644 --- a/frontend/src/lib/consultation-thinking-plan.ts +++ b/frontend/src/lib/consultation-thinking-plan.ts @@ -135,7 +135,7 @@ export const PERIOD_FIRST_RULE = "问一段时间(最近、今年、这段时 * The server overlay (answer_policy, deterministic_claims_forbidden_for) is * unchanged: it still forbids certainty, not the conditional reading. */ -export const INTERPRETATION_STATUS_RULE = "解读状态(references/interpretation_status_contract.md):parameter_sensitive = 有来源、已计算,但派别或参数会改变结果、外部验证未完成——照常解读,条件和所用参数写进「判断依据」,正文不讲参数和算法;user_history_verification_required = 已计算,需要用户的现实经历来核验——照常解读,写成待核验并给出可核验的问题;blocked = 没有输入或确实不能运行——只写缺什么,不产生结果性断言;not_available_yet = 本产品尚未接入的技法(如 Sarvatobhadra、Kota、Kalanala、Nakshatra Latta、部分稀有大运、问事盘)——在审计里写明,不假装用过,也不因此阻止其他技法的条件性分析。外部 oracle 未闭环、体系分歧、工具结果里的 reference_only、partial、observation_only,只降低断言强度、不等于不能解读;KP(宫头、Sign Lord、Star Lord、Sub Lord、significator)、行运和 Saham 可以有条件地作为依据,和整宫结果不一致时两面都写。服务端 answer_policy 禁止的仍然禁止:它限制的是确定性的说法,不是条件性的解读。"; +export const INTERPRETATION_STATUS_RULE = "解读状态(references/interpretation_status_contract.md):parameter_sensitive = 有来源、已计算,但派别或参数会改变结果、外部验证未完成——照常解读,条件和所用参数写进「判断依据」,正文不讲参数和算法;user_history_verification_required = 旧数据里的标记,不算缺口——按已有依据照常判断,不写待核验,不向用户要过往经历,也不因此降低结论(用户经历只在生时校正流程里使用);blocked = 没有输入或确实不能运行——只写缺什么,不产生结果性断言;not_available_yet = 本产品尚未接入的技法(如 Sarvatobhadra、Kota、Kalanala、Nakshatra Latta、部分稀有大运、问事盘)——在审计里写明,不假装用过,也不因此阻止其他技法的条件性分析。外部 oracle 未闭环、体系分歧、工具结果里的 reference_only、partial、observation_only,只降低断言强度、不等于不能解读;KP(宫头、Sign Lord、Star Lord、Sub Lord、significator)、行运和 Saham 可以有条件地作为依据,和整宫结果不一致时两面都写。服务端 answer_policy 禁止的仍然禁止:它限制的是确定性的说法,不是条件性的解读。"; export const PLAIN_SPEECH_RULE = `分析清单只用来想,不用来说:工具结果里的清单(受冲信号的编号与条数、「分三件事看」的层名、「旺弱不说明顺不顺」这类规则原句、「从某宫数起第几宫」这类数法)只决定你怎么判断,不写进回答,也不对用户解释规则;说给用户的是判断的结果,用生活里的话。依据用白话写进句子里,可以压缩语言、不必把信号逐条列全,但决定这句判断的反证不能省(占星师 2026-11-08 第五轮裁定 6:原值「说出来的依据最多两条,挑最能撑住这句话的」,删除「最多两条」)。${PLAIN_TERMS}`; diff --git a/frontend/src/lib/personal-report-generation.ts b/frontend/src/lib/personal-report-generation.ts index d1d4cf63c..c031c7a6b 100644 --- a/frontend/src/lib/personal-report-generation.ts +++ b/frontend/src/lib/personal-report-generation.ts @@ -2176,9 +2176,8 @@ export function buildReportEvidenceBundleV2( executedTechniqueRefs: uniqueMatched.map((receipt) => receipt.id), assertionLevel, timingBoundary, - verificationQuestions: assertionLevel === "parameter_sensitive" - ? ["建议结合真实经历核验本主题的部分证据结论"] - : [], + // BUG-2111: a normal reading never asks the reader to check it against their own past. + verificationQuestions: [], }); } @@ -2546,7 +2545,8 @@ const CLAIM_STATUS_RANK: Readonly> = { single_system_inference: 1, parameter_sensitive: 2, unclosed_divisional_chart: 3, - user_history_verification_required: 4, + // BUG-2111: missing user history is no penalty (astrologer 2026-11-21) — ranked with parameter_sensitive. + user_history_verification_required: 2, blocked: 5, }; diff --git a/frontend/src/lib/personal-report-reading-terms.ts b/frontend/src/lib/personal-report-reading-terms.ts index ccf0d21f9..c896e6547 100644 --- a/frontend/src/lib/personal-report-reading-terms.ts +++ b/frontend/src/lib/personal-report-reading-terms.ts @@ -74,6 +74,9 @@ export function withoutGlossedNames(text: string): string { const HEDGE = /(?:只|仅)(?:能)?(?:作|供|当|用作|作为)[^。,;,;::!?!?\s]{0,4}?参考/u; /** The 「只作参考」-type hedge in one sentence, or undefined. */ +/** BUG-2111 (astrologer 2026-11-21): 「未见阻断项」 is an internal state, never reader text. */ +const INTERNAL_STATE = /未见阻断项|no[ _]open[ _]blocker/iu; + export function readingHedge(text: string): string | undefined { - return HEDGE.exec(text)?.[0]; + return HEDGE.exec(text)?.[0] ?? INTERNAL_STATE.exec(text)?.[0]; } diff --git a/frontend/src/lib/personal-report-reading.ts b/frontend/src/lib/personal-report-reading.ts index 44761a0a2..417ffa36f 100644 --- a/frontend/src/lib/personal-report-reading.ts +++ b/frontend/src/lib/personal-report-reading.ts @@ -883,7 +883,7 @@ const REPAIR_INSTRUCTIONS: Partial> = { disclaimer: "将重复的边界说明合为一处。", banned_phrase: "删掉命中句里的确定性、医疗或其他超出报告范围的表达。", technique_name: "命中句写了技法或点位的名字。改成读者听得懂的话:说这颗星、这个点、这段时间对当事人意味着什么;时间直接写起止日期;行星、星座、宫位和日期照写。技法名只写在「依据与边界」。" + READER_WORDING_TAIL, - hedge_phrase: "命中句写了「只作参考」这类话。这一层依据用来看什么就直接写它看什么,用不上就不写;不换成别的说法再写一遍。" + READER_WORDING_TAIL, + hedge_phrase: "命中句写了「只作参考」这类话,或「未见阻断项」这类内部状态。这一层依据用来看什么就直接写它看什么,用不上就不写;不换成别的说法再写一遍。" + READER_WORDING_TAIL, strength_verdict: "命中句从行星力量推出了事实包里没有的判断(扛得住、稳得住、撑不撑得起来这类)。力量只照事实包写有没有达到最低线,不从力量推出对生活的判断;要说生活里会怎样,用宫位、主星和时间的依据。" + READER_WORDING_TAIL, dignity_mismatch: "把星曜状态改为事实包里的状态,无法核对时删除该说法。", false_missing_layer: "删除未计算的说法,按事实包写实际有无和状态。", diff --git a/frontend/src/lib/report-fact-table-display.ts b/frontend/src/lib/report-fact-table-display.ts index f4378b78d..adfafd2d3 100644 --- a/frontend/src/lib/report-fact-table-display.ts +++ b/frontend/src/lib/report-fact-table-display.ts @@ -6,7 +6,8 @@ export const FACT_STATUS_LABELS = { single_system_inference: "单一系统推断", parameter_sensitive: "参数敏感", unclosed_divisional_chart: "分盘尚未核验完整", - user_history_verification_required: "需要过往经历核对", + // BUG-2111: never a gap for the reader; normalised away in report-fact-tables.ts, neutral if it slips through. + user_history_verification_required: "按已有依据判断", blocked: "暂不可用", } as const; @@ -15,7 +16,7 @@ export const FACT_STATUS_LABELS_EN: Record>>> = { candidate_alignment_requires_event_history: { - zh: "两人各自都符合这一类事件的条件,是候选,需要用真实经历核对,不代表已经或一定发生", - en: "Both people individually meet the conditions for this event class. It is a candidate for checking against real history.", + // BUG-2111: no request to check against the couple's own history (astrologer 2026-11-21). + zh: "两人各自都符合这一类事件的条件,是候选,不代表已经或一定发生", + en: "Both people individually meet the conditions for this event class. It is a candidate, not something that has happened or must happen.", }, parallel_observation_only: { zh: "两人各有观察,但还不够判断", en: "There are observations for both people, with too little evidence to draw a conclusion" }, one_sided_or_incomplete: { zh: "单侧或证据不完整", en: "One-sided or incomplete evidence" }, diff --git a/frontend/src/lib/user-history-isolation.ts b/frontend/src/lib/user-history-isolation.ts new file mode 100644 index 000000000..3503aeb66 --- /dev/null +++ b/frontend/src/lib/user-history-isolation.ts @@ -0,0 +1,144 @@ +/** + * BUG-2111 (astrologer 2026-11-21, relayed by the owner): 「普通解盘与推运不得使用用户经历或反馈, + * 改变事件选择、权重、方向、计算参数或结论。只有用户明确进入生时校正流程时,才可在该流程内使用相关经历。 + * 独立公开案例的规则验证属于另外的研究流程,不等同于对当前用户进行经历校准。」 + * + * Her reference files still tell the model to calibrate on the user's past events + * (precision-reading-methodology.md 共识6 / §4.1–4.2 / §4.3 gate / grade A / Step 3; + * ai-reading-workflow-prompt.md §5.0.1–5.0.2 rows; transit-actionable-output-guide.md + * 要素3 line A). She will edit the upstream text herself. Until then the product cuts + * these passages out of every text that reaches a model in normal readings and + * forecasts. The files are not edited (owner rule: her skill runs in full; product + * differences live in announced product files — references/product_runtime_contract.md + * §AU). No rectification flow reads these files, so there is no exception to make. + * + * Passages are found by their own text, not by line numbers. When her edit removes a + * passage, its rule stops matching; tests/user-history-isolation-2111.test.ts then + * fails and names the rule, so the layer is dropped instead of silently lingering. + */ + +export const USER_HISTORY_ISOLATION_NOTE = + "(产品说明:此处原文讲用用户过往经历校准预测,普通解盘与推运不使用:不向用户索要过往经历,也不因为没有经历而降低结论或停止预测。用户经历只在生时校正流程里使用。)"; + +type BlockRule = Readonly<{ + id: string; + file: string; + /** The first line of the passage, matched after trimming. */ + start: string; + /** The passage stops before the first later line that starts with this (trimmed); "" = blank line. */ + end: string; + note: boolean; +}>; + +type LineRule = Readonly<{ + id: string; + file: string; + /** A line whose trimmed text starts with this is removed. */ + line: string; +}>; + +export type UserHistoryRule = BlockRule | LineRule; + +const METHODOLOGY = "references/precision-reading-methodology.md"; +const WORKFLOW = "references/ai-reading-workflow-prompt.md"; +const TRANSIT = "references/transit-actionable-output-guide.md"; +/** Her tables join two words with a middle dot; built here so no literal separator sits in our source. */ +const DOT = String.fromCharCode(0xb7); + +export const USER_HISTORY_RULES: readonly UserHistoryRule[] = [ + { id: "methodology-consensus-6", file: METHODOLOGY, start: "### 共识6:先验证过去,再预测未来", end: "---", note: true }, + { id: "methodology-4.1-4.2", file: METHODOLOGY, start: "### 4.1 为什么要验证过去", end: "### 4.3 ", note: true }, + { id: "methodology-4.3-gate", file: METHODOLOGY, line: "**只有在倒推验证通过后,才进入正推预测。**" }, + { id: "methodology-grade-a", file: METHODOLOGY, line: "| **A-已验证** | 经过用户已知事件倒推验证的结论" }, + { id: "methodology-step-3", file: METHODOLOGY, start: "Step 3: 倒推验证", end: "", note: false }, + { id: "workflow-grade-a", file: WORKFLOW, line: "| **A-已验证** | ✅ [A] | 经过用户已知事件倒推验证的结论" }, + { id: "workflow-verified", file: WORKFLOW, line: `| 已验证 | \`[已验证${DOT}N/N事件吻合]\`` }, + { id: "workflow-unverified", file: WORKFLOW, line: "| 未经验证 | `[未经验证]` | 用户未提供过往事件" }, + { id: "workflow-calibration-failed", file: WORKFLOW, line: `| 验证失败 | \`[校准异常${DOT}需重新评估]\`` }, + { id: "transit-grade-a", file: TRANSIT, line: "✅ [A] 已验证(倒推历史事件命中)" }, +]; + +/** Cheap pre-check: a text can only need isolating if it carries one of these. */ +const MARKERS = [...new Set(USER_HISTORY_RULES.map((rule) => ("line" in rule ? rule.line : rule.start)))]; + +export function mayNeedUserHistoryIsolation(text: string): boolean { + return MARKERS.some((marker) => text.includes(marker)); +} + +export type UserHistoryIsolation = Readonly<{ text: string; applied: readonly string[] }>; + +/** Cut every matching passage out of `text`. Text without a marker comes back unchanged. */ +export function isolateUserHistoryCalibration(text: string): UserHistoryIsolation { + if (!mayNeedUserHistoryIsolation(text)) return { text, applied: [] }; + let lines = text.split("\n"); + const applied: string[] = []; + for (const rule of USER_HISTORY_RULES) { + if ("line" in rule) { + const kept = lines.filter((line) => !line.trim().startsWith(rule.line)); + if (kept.length !== lines.length) { + applied.push(rule.id); + lines = kept; + } + continue; + } + const out: string[] = []; + let hit = false; + for (let index = 0; index < lines.length; index += 1) { + if (lines[index].trim() !== rule.start && !lines[index].trim().startsWith(rule.start)) { + out.push(lines[index]); + continue; + } + let stop = index + 1; + while (stop < lines.length) { + const trimmed = lines[stop].trim(); + if (rule.end === "" ? trimmed === "" : trimmed.startsWith(rule.end)) break; + stop += 1; + } + if (rule.note) out.push(USER_HISTORY_ISOLATION_NOTE, ""); + hit = true; + index = stop - 1; + } + if (hit) { + applied.push(rule.id); + lines = out; + } + } + return { text: lines.join("\n"), applied }; +} + +/** + * Apply the isolation to every string inside a value (a model prompt, a tool result), + * returning the same object when nothing matched so callers can skip a rewrite. + */ +export function isolateUserHistoryDeep(value: T): { value: T; applied: string[] } { + const applied: string[] = []; + const visit = (node: unknown): unknown => { + if (typeof node === "string") { + if (!mayNeedUserHistoryIsolation(node)) return node; + const result = isolateUserHistoryCalibration(node); + applied.push(...result.applied); + return result.text; + } + if (Array.isArray(node)) { + let changed = false; + const next = node.map((item) => { + const mapped = visit(item); + if (mapped !== item) changed = true; + return mapped; + }); + return changed ? next : node; + } + if (node && typeof node === "object" && Object.getPrototypeOf(node) === Object.prototype) { + let changed = false; + const next: Record = {}; + for (const [key, item] of Object.entries(node as Record)) { + const mapped = visit(item); + if (mapped !== item) changed = true; + next[key] = mapped; + } + return changed ? next : node; + } + return node; + }; + return { value: visit(value) as T, applied: [...new Set(applied)] }; +} diff --git a/frontend/src/mastra/personal-report.ts b/frontend/src/mastra/personal-report.ts index fcfbf30ec..3e9f4c8b6 100644 --- a/frontend/src/mastra/personal-report.ts +++ b/frontend/src/mastra/personal-report.ts @@ -29,12 +29,14 @@ export type { * stacks must never reach this agent. */ +// BUG-2111 (astrologer 2026-11-21): the writer is no longer offered user_history_verification_required — +// missing user history is not a gap in a normal reading. Stored reports still validate through +// lib/personal-report-contract.ts CLAIM_STATUSES, which keeps the value. export const claimStatusSchema = z.enum([ "multi_system_consensus", "single_system_inference", "parameter_sensitive", "unclosed_divisional_chart", - "user_history_verification_required", "blocked", ]); diff --git a/frontend/src/mastra/product-voice.ts b/frontend/src/mastra/product-voice.ts index ad30a53a6..86f112a36 100644 --- a/frontend/src/mastra/product-voice.ts +++ b/frontend/src/mastra/product-voice.ts @@ -132,8 +132,8 @@ export const professionalConversationVoice = `VISIBLE VOICE ## 实际执行的技法与原始数据:先 D1(上升与度数、相关宫的星座和宫内星、宫主落宫、相关行星的度数、尊贵和燃烧);再功能吉凶——按当前上升写出功能吉星、功能凶星、功能中性星和 Yogakaraka,自然吉凶与功能吉凶冲突的星写出冲突来源并降低置信度;再力量(Shadbala 达标与否、Bhava Bala、SAV / BAV);再 Yoga(只写卡上 yogas_detail 和 yogas.hits 列出的,状态照卡上原词写,写条件和是否被当前 Dasha 激活);再 Jaimini(AK、AmK、DK、Karakamsha、AL、UL、A7、A10 等按领域取)。 ## 相关分盘:按问题域取——事业 D10 + A10 + AmK;财富 D2、D11、A2、A11(收入来自工作时加 D10、A10),分开「能不能挣到」和「能不能留下」;婚恋 D9 + UL + A7 + DK;子女 D7;父母 D12;学习 D24;健康 D6、D8、D30;迁居 D4、D12。与本题无关的分盘不列。 ## 多套 Dasha 与行运:Vimshottari 大运、小运、细运的起止,Narayana 主运、子运的起止,相关的 Jaimini dasha(Chara 等;卡上没有时用 read-consultation-evidence 补取),木星、土星、罗睺、计都的行运与双重行运,年运(Tajika 年盘、Muntha、年主;卡上有才写)。几套方向不同时,两面都写出来:各自的机制和让它成立的条件,不只挑一套。时间节点分开写:接触、启动、确认、落地、责任考验。 -## 证据等级:每项实质判断标 A / B / C(router §10:A = 本命承诺 + 相关分盘 + Dasha + 行运 + 至少一层核验,个人预测最好有历史回测;B = 三套主要体系支持,但缺一层或时间宽;C = 只有一两个指标),并标解读状态。每个等级后面写它按的规则和满足、缺少的项,例如「B(router §10:本命、分盘、Dasha 支持,行运时间宽)」,读者能照着复核。本轮不写 A(A 要求的历史回测与核验层本产品没有),也不在正文解释这一点。 -## 冲突与阻塞项:体系之间、自然与功能吉凶之间、D1 与分盘之间、两套 Dasha 之间的冲突;blocked(写缺什么输入)、not_available_yet、parameter_sensitive、user_history_verification_required 的项目各自写明。 +## 证据等级:每项实质判断标 A / B / C(router §10:A = 本命承诺 + 相关分盘 + Dasha + 行运 + 至少一层核验;B = 三套主要体系支持,但缺一层或时间宽;C = 只有一两个指标),并标解读状态。每个等级后面写它按的规则和满足、缺少的项,例如「B(router §10:本命、分盘、Dasha 支持,行运时间宽)」,读者能照着复核。本轮不写 A(A 要求的核验层本产品没有),也不在正文解释这一点。等级只看盘面依据够不够,不看用户有没有讲过往经历;普通解读不向用户索要过往经历来定等级。 +## 冲突与阻塞项:体系之间、自然与功能吉凶之间、D1 与分盘之间、两套 Dasha 之间的冲突;blocked(写缺什么输入)、not_available_yet、parameter_sensitive 的项目各自写明。工具结果里出现 user_history_verification_required 时不算缺口:按已有依据照常判断,不写待核验,不向用户要过往经历。 ## 结论与行动建议:展开开头的结论;按用户的问法分段,每条结论回指上面的证据和等级;问时间的按时间段写起止和做什么;给两三件具体能做的事。 ## Technique Audit Table:Markdown 表格,五列 Technique | Calculation | Interpretation | Key result | Effect on confidence;每行两个状态分开写,不合成一个值(占星师 2026-11-08 第五轮裁定 7,她的 interpretation_status_contract)。Calculation(计算状态)只用 executed、partial、blocked、not_applicable:算没算出来;not_available_yet 写 blocked,原因写在 Key result。Interpretation(解读状态)只用 primary、supporting、conditional、observation_only、blocked、not_used:这一层的结果怎样参与本题的判断——primary 主要依据,supporting 辅助依据,conditional 有条件(参数敏感、方法或版本有分歧),observation_only 只作观察,blocked 不作结果性判断,not_used 本题不参与。能算但方法有分歧 = Calculation executed + Interpretation conditional,不写 blocked:卡上 Narayana 的 method_status.chat_use 为 single_system_candidate 就是这一档,Effect 写不据此强化应期;Calculation 写 blocked 只给没有结果的层(起运、双宫主或年数定不出,chat_use 为 no_judgment;缺输入;未接入),这时 Interpretation 也是 blocked。卡上标 parameter_sensitive 的(KP、Tajika 年盘、Shadbala 达标与否)写 executed + conditional;not_applicable 写 not_applicable + not_used。照卡上每个领域的 route 表列:on_card 且本题用到的写 executed,not_applicable、blocked、not_available_yet 照它写并写明缺什么。至少列:D1、D9、本题相关分盘、Functional Benefic/Malefic(Key result 写关键功能吉星和功能凶星,并写所用的功能吉凶表 profile)、Vimshottari、Narayana、Jaimini / Chara Dasha、Shadbala、Ashtakavarga、Yoga、Transit、Tajika / Varshaphala、KP / Sub-lord;router §8 里和本题相关的其他行也列。不写外部验证、联网、MEVG 或真实案例校准的行(产品 2026-11-08 决定:这些只记在后台回执),也不写「已验证」「MEVG 通过」一类的总评。 diff --git a/frontend/src/mastra/report-reading-skill.ts b/frontend/src/mastra/report-reading-skill.ts index c8e116a0f..794c3c16a 100644 --- a/frontend/src/mastra/report-reading-skill.ts +++ b/frontend/src/mastra/report-reading-skill.ts @@ -8,6 +8,7 @@ import { PLAIN_TERMS } from "../lib/consultation-thinking-plan.ts"; import type { ReportReadingPrompt } from "../lib/personal-report-reading.ts"; import { jyotishSkillPackage } from "./skill-binding.ts"; import { SYNASTRY_PLAIN_TIME_RULE } from "./synastry-report-chat.ts"; +import { isolateUserHistoryCalibration } from "../lib/user-history-isolation.ts"; /** * #192 (owner 2026-11-11 decision 2; astrologer 2026-11-12, her repo 8aaf6099 @@ -54,7 +55,12 @@ export function reportReadingPrompt(): ReportReadingPrompt { const skill = read("SKILL.md"); const body = skillBody(skill.text); if (!body) throw new Error(`Skill ${jyotishSkillPackage.name} has no method body for the report reading`); - const references = REPORT_READING_REFERENCES.map((path) => ({ path, ...read(path) })); + // BUG-2111: the hash stays the file's own; the text the model reads has the + // user-history calibration passages cut out (lib/user-history-isolation.ts). + const references = REPORT_READING_REFERENCES.map((path) => { + const file = read(path); + return { path, sha256: file.sha256, text: isolateUserHistoryCalibration(file.text).text }; + }); const contract = read(PRODUCT_RUNTIME_CONTRACT_FILE); const sourceNotices = productContractBlock("report-reading-source-notices", {}, contract.text); // BUG-2098: the reading's sections for ordinary readers take the chat's own plain-terms and time rules. diff --git a/frontend/src/mastra/skill-binding.ts b/frontend/src/mastra/skill-binding.ts index 85b615b56..677456b85 100644 --- a/frontend/src/mastra/skill-binding.ts +++ b/frontend/src/mastra/skill-binding.ts @@ -6,6 +6,7 @@ import { resolveLiveJyotishSkillRuntimePath, } from "../lib/skill-package-registry.ts"; import { productContractBlock } from "../lib/product-runtime-contract.ts"; +import { isolateUserHistoryDeep } from "../lib/user-history-isolation.ts"; const skill = resolveLiveJyotishSkill(); @@ -178,7 +179,10 @@ export const jyotishSkillMethodBlock = methodBlock(NATAL_REPORT_SKELETON); * that bind this skill share this object, so remembering that one of them had * its method would stop the others from ever being asked. */ -export const jyotishSkillBoundProcessor: Processor & { processInputStep: NonNullable } = { +export const jyotishSkillBoundProcessor: Processor & { + processInputStep: NonNullable; + processLLMRequest: NonNullable; +} = { id: "jyotish-skill-bound", name: "Jyotish Skill Bound", providesSkillDiscovery: "on-demand", @@ -189,6 +193,16 @@ export const jyotishSkillBoundProcessor: Processor & { processInputStep: NonNull abort(`Jyotish skill method is not bound into the system prompt for ${skill.name}`); } }, + /** + * BUG-2111: `skill_read` hands the model any reference file whole, so the + * passages that calibrate on the user's past events (lib/user-history-isolation.ts) + * are cut out of the prompt here, on every call, before it reaches the provider. + * Nothing is persisted: the stored tool result keeps the file as read. + */ + processLLMRequest({ prompt }) { + const { value, applied } = isolateUserHistoryDeep(prompt); + return applied.length > 0 ? { prompt: value } : undefined; + }, }; /** diff --git a/frontend/tests/consult-chat-reopt-20261107.test.ts b/frontend/tests/consult-chat-reopt-20261107.test.ts index 325b96f3f..b595880d6 100644 --- a/frontend/tests/consult-chat-reopt-20261107.test.ts +++ b/frontend/tests/consult-chat-reopt-20261107.test.ts @@ -18,6 +18,7 @@ import { import { resolveLiveJyotishSkill } from "../src/lib/skill-package-registry.ts"; import { consultationWorkflowResponseSchema, toAgentConsultationContext, toModelOutput } from "../src/mastra/consultation-workflow.ts"; import { createConsultationPlan } from "../src/lib/consultation-plan.ts"; +import { isolateUserHistoryCalibration } from "../src/lib/user-history-isolation.ts"; type Json = Record; const golden = JSON.parse(readFileSync(new URL("./fixtures/consult-evidence-card-golden.json", import.meta.url), "utf8")) as { @@ -67,7 +68,9 @@ test("BUG-1726: her files travel in her three layers, verbatim where whole", () const timed = consultationMethodologyForDomains(["career"], { question: "我什么时候升职" })!; const transit = timed.delivery.find((row) => row.source === "references/transit-actionable-output-guide.md"); assert.deepEqual(transit && { layer: transit.layer }, { layer: "per_question" }); - assert.equal(timed.sections.find((s) => s.source === "references/transit-actionable-output-guide.md")!.text, herFile("references/transit-actionable-output-guide.md")); + // 原值: 与 herFile(...) 逐字相等;新值: 与去掉用户经历校准段落后的原文逐字相等; + // 原因: BUG-2111(占星师 2026-11-21:普通解读不用用户经历;§AU 产品隔离,原文件不改) + assert.equal(timed.sections.find((s) => s.source === "references/transit-actionable-output-guide.md")!.text, isolateUserHistoryCalibration(herFile("references/transit-actionable-output-guide.md")).text); // Health has no own adjudicator: the generic skeleton travels whole, and the §5–§6 excerpt is not repeated. const health = consultationMethodologyForDomains(["health"])!; assert.equal(health.sections.find((s) => s.source === "references/event_judgment_skeleton.md")!.text, herFile("references/event_judgment_skeleton.md")); diff --git a/frontend/tests/personal-report-api.test.ts b/frontend/tests/personal-report-api.test.ts index edc5eb3f5..778f30030 100644 --- a/frontend/tests/personal-report-api.test.ts +++ b/frontend/tests/personal-report-api.test.ts @@ -132,7 +132,7 @@ function agentOutputForBundle( narrative: claim.conclusion, actions: [], caveats: [], - claimStatus: claim.assertionLevel, + claimStatus: writerStatus(claim.assertionLevel), evidenceRefs: [...section.evidenceRefs], }; }), @@ -1540,3 +1540,9 @@ test("core create: a billed report has no daily or hourly cap (BUG-1641)", async assert.equal(counted, 0); assert.equal(billing.reservations, 1); }); + +// Defined last so the privacy allow-list line numbers above stay put. +// BUG-2111: the writer is no longer offered user_history_verification_required; a card that still +// carries it (old data) is written as parameter_sensitive, the rank it now shares. +const writerStatus = (level: T) => + (level === "user_history_verification_required" ? "parameter_sensitive" : level) as Exclude; diff --git a/frontend/tests/personal-report-generation-v2.test.ts b/frontend/tests/personal-report-generation-v2.test.ts index a1b27c68a..49713d778 100644 --- a/frontend/tests/personal-report-generation-v2.test.ts +++ b/frontend/tests/personal-report-generation-v2.test.ts @@ -28,6 +28,11 @@ import type { ReportAgentPort, } from "../src/mastra/personal-report.ts"; +// BUG-2111: the writer is no longer offered user_history_verification_required; a card that still +// carries it (old data) is written as parameter_sensitive, the rank it now shares. +const writerStatus = (level: T) => + (level === "user_history_verification_required" ? "parameter_sensitive" : level) as Exclude; + const REPORT_ID = "22222222-2222-4222-8222-222222222222"; const GENERATED_AT = "2026-08-14T00:00:00.000Z"; const SIGNS = [ @@ -206,7 +211,7 @@ function writerOutputFor( narrative: card.conclusion, actions: [`围绕${card.section}记录可验证的现实反馈`], caveats: ["该结论不得脱离所列证据引用。"], - claimStatus: card.assertionLevel, + claimStatus: writerStatus(card.assertionLevel), evidenceRefs: [...section.evidenceRefs], }; }), @@ -844,7 +849,7 @@ function sectionPayloadFor( narrative: card.conclusion, actions: [`围绕${card.section}记录可验证的现实反馈`], caveats: ["该结论不得脱离所列证据引用。"], - claimStatus: card.assertionLevel, + claimStatus: writerStatus(card.assertionLevel), evidenceRefs: [...section.evidenceRefs], }; } diff --git a/frontend/tests/report-fact-tables.test.ts b/frontend/tests/report-fact-tables.test.ts index 18429c9cb..5183a9d86 100644 --- a/frontend/tests/report-fact-tables.test.ts +++ b/frontend/tests/report-fact-tables.test.ts @@ -101,7 +101,9 @@ test("density golden: Sade Sati reports missing cycle dates without inventing th test("density facts preserve declared limitations and reject malformed provenance", () => { const altered = structuredClone(packet); altered.worksheets.advanced_systems.yogas_doshas.sade_sati.claimStatus = "user_history_verification_required"; - assert.equal(assembleReportFactTables(altered).find(t => t.id === "sade_sati")?.claimStatus, "user_history_verification_required"); + // 原值: 原样保留 "user_history_verification_required";新值: 按未声明状态读成 "parameter_sensitive"; + // 原因: BUG-2111(占星师 2026-11-21:缺用户经历不算缺口,读者不看到「需要过往经历核对」) + assert.equal(assembleReportFactTables(altered).find(t => t.id === "sade_sati")?.claimStatus, "parameter_sensitive"); const tables = assembleReportFactTables(packet); tables[0].rows[0].sourcePath = tables[0].sourcePath + "_wrong.lord"; assert.equal(reportFactTablesSchema.safeParse(tables).success, false); diff --git a/frontend/tests/synastry-exploratory-2014.test.tsx b/frontend/tests/synastry-exploratory-2014.test.tsx index f8ee4b460..4ac885f77 100644 --- a/frontend/tests/synastry-exploratory-2014.test.tsx +++ b/frontend/tests/synastry-exploratory-2014.test.tsx @@ -39,7 +39,9 @@ test("BUG-2014: a formal candidate with scan periods is a body row; a scan-only const saved = eventClassRows(dossier, names, "zh"); const body = readerSynastryEvents(saved, "zh", names); assert.deepEqual(body.map((row) => row.title), ["心动"]); - assert.match(body[0].formal!, /^正式核对:两人各自都符合这一类事件的条件,是候选,需要用真实经历核对/); + // 原值: /…是候选,需要用真实经历核对/;新值不再要求用两人的经历核对;原因: BUG-2111(占星师 2026-11-21:普通解读不用用户经历) + assert.match(body[0].formal!, /^正式核对:两人各自都符合这一类事件的条件,是候选,不代表已经或一定发生/); + assert.doesNotMatch(body[0].formal!, /真实经历|过往经历/); assert.equal(body[0].research, "研究扫描:两人同时有相关运限:2027 年 3 月至 9 月"); assert.deepEqual(body[0].people, []); const scanned = exploratorySynastryEvents(saved, "zh", names); diff --git a/frontend/tests/user-history-isolation-2111.test.ts b/frontend/tests/user-history-isolation-2111.test.ts new file mode 100644 index 000000000..f57f64ade --- /dev/null +++ b/frontend/tests/user-history-isolation-2111.test.ts @@ -0,0 +1,133 @@ +/** + * BUG-2111 (astrologer 2026-11-21): normal readings and forecasts never use the user's own past + * events. Her reference files still say so in places; the product cuts those passages out of every + * text sent to a model (references/product_runtime_contract.md §AU) until she edits them upstream. + */ +import assert from "node:assert/strict"; +import { readFileSync } from "node:fs"; +import { resolve } from "node:path"; +import test from "node:test"; + +import { + USER_HISTORY_ISOLATION_NOTE, + USER_HISTORY_RULES, + isolateUserHistoryCalibration, + isolateUserHistoryDeep, +} from "../src/lib/user-history-isolation.ts"; +import { reportReadingPrompt } from "../src/mastra/report-reading-skill.ts"; +import { jyotishSkillBoundProcessor } from "../src/mastra/skill-binding.ts"; +import { consultationMethodologyForDomains } from "../src/lib/consultation-methodology.ts"; +import { INTERPRETATION_STATUS_RULE } from "../src/lib/consultation-thinking-plan.ts"; +import { annualReplyWording } from "../src/lib/annual-question-check.ts"; +import { readingHedge } from "../src/lib/personal-report-reading-terms.ts"; +import { FACT_STATUS_LABELS, FACT_STATUS_LABELS_EN } from "../src/lib/report-fact-table-display.ts"; +import { EVIDENCE_LABELS } from "../src/lib/synastry-full-sections.ts"; +import { claimStatusSchema as writerClaimStatus } from "../src/mastra/personal-report.ts"; +import { CLAIM_STATUSES } from "../src/lib/personal-report-contract.ts"; + +const ROOT = resolve(import.meta.dirname, "../.."); +const file = (path: string) => readFileSync(resolve(ROOT, path), "utf8").replace(/\r\n/g, "\n"); + +/** Phrases that only the isolated passages carry; none may reach a model after isolation. */ +const FORBIDDEN: Readonly> = { + "references/precision-reading-methodology.md": [ + "先验证过去,再预测未来", "用户的真实经历是最好的校准数据", "倒推验证协议", "收集已知事件", + "必须停止预测,先解决校准问题", "只有在倒推验证通过后", "经过用户已知事件倒推验证", "用用户已知的3-5件重大事件", + "在预测未来之前,必须先用已知的过去事件做验证校准", + ], + "references/ai-reading-workflow-prompt.md": ["经过用户已知事件倒推验证", "用户未提供过往事件", "[校准异常·需重新评估]", "[已验证·N/N事件吻合]"], + "references/transit-actionable-output-guide.md": ["已验证(倒推历史事件命中)"], +}; + +test("every isolation rule still matches its upstream passage (fails when she edits the file: drop that rule)", () => { + for (const rule of USER_HISTORY_RULES) { + const { applied } = isolateUserHistoryCalibration(file(rule.file)); + assert.ok( + applied.includes(rule.id), + `rule ${rule.id} no longer matches ${rule.file}: if the astrologer removed this passage upstream, delete the rule (and the layer when none is left)`, + ); + } +}); + +test("after isolation no user-history calibration phrase is left, and the rest of the method is intact", () => { + for (const [path, phrases] of Object.entries(FORBIDDEN)) { + const before = file(path); + const after = isolateUserHistoryCalibration(before).text; + for (const phrase of phrases) { + assert.ok(before.includes(phrase), `${path} should carry 「${phrase}」 upstream (update FORBIDDEN if she changed it)`); + assert.ok(!after.includes(phrase), `${path} still sends 「${phrase}」`); + } + } + const methodology = isolateUserHistoryCalibration(file("references/precision-reading-methodology.md")).text; + for (const kept of ["## 二、PACDARES 本命盘分析框架", "### 4.3 正推预测协议", "**步骤1:确认承诺**", "| **B-强推断** |", "| **C-假设** |", "Step 4: 推运定位"]) { + assert.ok(methodology.includes(kept), `kept text missing: ${kept}`); + } + assert.equal(methodology.split(USER_HISTORY_ISOLATION_NOTE).length - 1, 2); + const unrelated = "## 标题\n\n普通文字,不含任何需要隔离的段落。\n"; + assert.equal(isolateUserHistoryCalibration(unrelated).text, unrelated); +}); + +test("the personal report reading quotes the methodology without the calibration passages, hash unchanged", () => { + const prompt = reportReadingPrompt(); + for (const phrase of FORBIDDEN["references/precision-reading-methodology.md"]) { + assert.ok(!prompt.system.includes(phrase), `reading prompt still sends 「${phrase}」`); + } + assert.ok(prompt.system.includes(USER_HISTORY_ISOLATION_NOTE)); + const loaded = prompt.skill.loaded.find((entry) => entry.path === "references/precision-reading-methodology.md"); + assert.ok(loaded && /^[0-9a-f]{64}$/.test(loaded.sha256), "the file's own hash is still recorded"); +}); + +test("the chat's per-call processor cuts the passages out of a skill_read result and leaves other prompts alone", async () => { + const raw = file("references/precision-reading-methodology.md"); + const prompt = [ + { role: "system", content: "system text" }, + { role: "tool", content: [{ type: "tool-result", toolCallId: "1", toolName: "skill_read", output: { type: "text", value: raw } }] }, + ]; + const args = { prompt } as unknown as Parameters[0]; + const result = await jyotishSkillBoundProcessor.processLLMRequest(args); + const sent = JSON.stringify(result && "prompt" in result ? result.prompt : prompt); + for (const phrase of FORBIDDEN["references/precision-reading-methodology.md"]) assert.ok(!sent.includes(phrase), phrase); + assert.ok(sent.includes("PACDARES")); + assert.ok(raw.includes("倒推验证协议"), "the stored tool result object is not mutated"); + + const plain = [{ role: "user", content: [{ type: "text", text: "我的事业怎么样?" }] }]; + const untouched = await jyotishSkillBoundProcessor.processLLMRequest({ prompt: plain } as unknown as typeof args); + assert.equal(untouched, undefined); + const deep = isolateUserHistoryDeep(plain); + assert.equal(deep.value, plain); +}); + +test("the chat's quoted transit guide no longer carries grade A by back-calculated past events", () => { + const timing = consultationMethodologyForDomains(["timing"])!; + const transit = timing.sections.find((section) => section.source === "references/transit-actionable-output-guide.md"); + assert.ok(transit, "the transit guide is quoted for a timing question"); + assert.ok(!transit.text.includes("已验证(倒推历史事件命中)")); + assert.ok(transit.text.includes("[B] 高概率")); +}); + +test("grades and statuses no longer depend on the user's past events", () => { + const voice = file("frontend/src/mastra/product-voice.ts"); + assert.ok(!voice.includes("历史回测"), "grade A no longer asks for a historical back-test"); + assert.ok(!/user_history_verification_required 的项目各自写明/.test(voice)); + assert.ok(!INTERPRETATION_STATUS_RULE.includes("可核验的问题")); + assert.ok(!INTERPRETATION_STATUS_RULE.includes("需要用户的现实经历来核验")); + assert.match(INTERPRETATION_STATUS_RULE, /user_history_verification_required = 旧数据里的标记,不算缺口/); + assert.ok(!(EVIDENCE_LABELS as readonly string[]).includes("user_history_verification_required")); + const synastrySections = file("frontend/src/lib/synastry-full-sections.ts"); + assert.ok(!/真实经历核对|need real history/.test(synastrySections), "no synastry section asks which judgments need the couple's history"); + assert.ok(!(writerClaimStatus.options as readonly string[]).includes("user_history_verification_required"), "the report writer is not offered it"); + assert.ok((CLAIM_STATUSES as readonly string[]).includes("user_history_verification_required"), "stored reports still validate"); + assert.ok(!/过往经历|past events/i.test(FACT_STATUS_LABELS.user_history_verification_required + FACT_STATUS_LABELS_EN.user_history_verification_required)); + const generation = file("frontend/src/lib/personal-report-generation.ts"); + assert.ok(!generation.includes("建议结合真实经历核验")); + assert.match(generation, /user_history_verification_required: 2,/); + const tables = file("frontend/src/lib/report-fact-tables.ts"); + assert.ok(!/declared === "user_history_verification_required"/.test(tables), "the fact table reads it like an undeclared status"); +}); + +test("「未见阻断项」 is an internal state: the reading and question checks catch it", () => { + assert.equal(readingHedge("这一块未见阻断项,可以放心。"), "未见阻断项"); + assert.ok(readingHedge("status no_open_blocker")); + assert.ok(annualReplyWording("这一项未见阻断项。").some((note) => note.includes("未见阻断项"))); + assert.equal(readingHedge("这一年工作上会更忙。"), undefined); +}); diff --git a/references/product_runtime_contract.md b/references/product_runtime_contract.md index 1d72d0107..d0c6d3c53 100644 --- a/references/product_runtime_contract.md +++ b/references/product_runtime_contract.md @@ -800,3 +800,14 @@ evidence_card.backstage records, for the backend only, that global web evidence - 个人报告:写好的日期后面不留空格;`not_routed` 的「没有路由」写成「已放在××部分使用」或「没有用到」;披露标签单独成句时写「本报告没有评估出生时间对各类判断的影响」;「现在这几年」第一句用日常话说阶段(规则 R-03);报告开头只在「重点」块保留一句「不是确定会发生的事」。 - 合盘「技术依据摘要」(只改显示):「档案时刻 / 候选时刻」显示为「填写的时间 / 试算的时间」;「参数敏感 / 只作观察」显示为「结果会随算法设置变化 / 只作参考」;表头用两个人的名字;同一句说明只留一次。 - 没有改:年运开始页「具体会发生什么事,星盘只能推测,报告里会写成『可能』,不当成定论。」(2026-11-01 的固定文案,#153);页首「本报告由 AI 生成,仅供参考。」 + +## AU. 普通解盘与推运不使用用户经历校准(占星师 2026-11-21 审核,BUG-2111;临时隔离) + +依据:占星师 2026-11-21 审核原话:「普通解盘与推运不得使用用户经历或反馈,改变事件选择、权重、方向、计算参数或结论。只有用户明确进入生时校正流程时,才可在该流程内使用相关经历。独立公开案例的规则验证属于另外的研究流程,不等同于对当前用户进行经历校准。」她会自己改上游原文;在那之前,产品把相关段落从送给模型的文字里隔离出去,参考文件本身不改。 + +- 隔离的段落(按原文内容匹配,不按行号):`precision-reading-methodology.md` 的「共识6:先验证过去,再预测未来」、§4.1–4.2(为什么要验证过去、倒推验证协议)、§4.3 开头「只有在倒推验证通过后,才进入正推预测」、§5.1 表里「A-已验证」一行、§八 的「Step 3: 倒推验证」;`ai-reading-workflow-prompt.md` §5.0.1 的「A-已验证」一行和 §5.0.2 的「已验证 / 未经验证 / 验证失败」三行;`transit-actionable-output-guide.md` 要素3 的「[A] 已验证(倒推历史事件命中)」一行。整段处换成一句产品说明,单行直接去掉。 +- 作用范围:个人报告解读(整段引用的参考文件)、问答(方法摘录和 `skill_read` 读到的任何参考文件,在每次发给模型前处理)。生时纠正流程不读这些文件,不受影响。 +- 等级:A 不再以「历史回测」为条件,等级只看盘面依据够不够;普通解读不向用户索要过往经历来定等级,也不因为没有经历而降低结论或停止预测。 +- `user_history_verification_required`:旧数据里仍可能出现。前端不把它当缺口:不显示「需要过往经历核对」,事实表按未声明状态处理,个人报告的状态排序与 `parameter_sensitive` 同级,提示词里不要求「写成待核验并给出可核验的问题」;合盘写作不再提供这个标签。真正缺输入、缺规则或方法冲突时,用已有的具体状态并写清缺什么(Python 一侧,BUG-2110)。 +- 「未见阻断项」(`no_open_blocker`)只是内部状态,不等于证据充分,不进读者文字;解读和报告内提问的回答出现时走同一次修复或改写。 +- 临时性:她改完上游后,这些段落不再出现,`tests/user-history-isolation-2111.test.ts` 会指出哪条规则不再匹配,届时删掉对应规则,全部不匹配时删掉这一层。