From 887b40df43095581519ab13f4ea71e8c873fd22d Mon Sep 17 00:00:00 2001
From: Jesse_Chen <chen.yue@mevixaga.com>
Date: Tue, 21 Jul 2026 05:23:51 +0800
Subject: [PATCH] fix: persist rectification evidence corrections

---
 ...onversational-birth-time-rectification.tsx |  29 +++-
 .../hooks/use-conversational-rectification.ts |  54 ++++++-
 .../conversational-rectification/contracts.ts |   3 +
 .../evidence-extractor.ts                     |   8 +-
 .../orchestrator.ts                           |  94 +++++++++++-
 .../persistence-contracts.ts                  |  14 +-
 ...00_conversational_rectification_schema.sql |  42 +++++-
 ...nversational_rectification_transitions.sql |  45 +++++-
 .../conversational-evidence-extractor.test.ts |  20 ++-
 ...ersational-rectification-component.test.ts |  61 +++++++-
 ...ersational-rectification-contracts.test.ts |  35 ++++-
 ...rsational-rectification-controller.test.ts | 106 ++++++++++++++
 ...ational-rectification-orchestrator.test.ts | 131 +++++++++++++++++
 ...conversational-rectification-store.test.ts |  24 ++-
 ...t_conversational_rectification_contract.py |  24 +++
 ...sational_rectification_postgres_runtime.py | 138 ++++++++++++++++++
 16 files changed, 801 insertions(+), 27 deletions(-)

diff --git a/frontend/src/components/conversational-birth-time-rectification.tsx b/frontend/src/components/conversational-birth-time-rectification.tsx
index 17186382..b433f30f 100644
--- a/frontend/src/components/conversational-birth-time-rectification.tsx
+++ b/frontend/src/components/conversational-birth-time-rectification.tsx
@@ -243,14 +243,33 @@ export function ConversationalRectificationSurface({
           submit();
         }}
       >
+        {controller.correctionTarget && (
+          <div className="conversational-correction-target" role="status">
+            <p>
+              正在更正：<strong>{controller.correctionTarget.dateLabel} · {controller.correctionTarget.summary}</strong>
+            </p>
+            <span>提交后会保留原记录用于审计，但候选评分只使用更正后的有效证据。</span>
+            <button
+              className="button-secondary"
+              disabled={controller.pending}
+              type="button"
+              onClick={() => controller.cancelEvidenceCorrection()}
+            >
+              取消更正
+            </button>
+          </div>
+        )}
         <label htmlFor="conversational-rectification-answer">
-          补充或更正真实经历
+          {controller.correctionTarget ? "填写更正后的真实经历" : "补充真实经历"}
           <span>大概年份也可以；不确定的部分请直接说不确定。</span>
         </label>
         <textarea
           id="conversational-rectification-answer"
           disabled={controller.pending || !canAnswer}
           maxLength={4_000}
+          placeholder={controller.correctionTarget
+            ? "例如：其实是 2020 年 11 月离职"
+            : "请描述一件已经发生的事，并尽量写明年份或月份"}
           ref={composer}
           rows={4}
           value={controller.draft}
@@ -283,13 +302,17 @@ export function ConversationalRectificationSurface({
           <ul>
             {turn.evidenceRecap.map((entry) => (
               <li key={entry.id}>
-                <div><time>{entry.dateLabel}</time><p>{entry.summary}</p></div>
+                <div>
+                  <time>{entry.dateLabel}</time>
+                  <p>{entry.summary}</p>
+                  {entry.isCorrection && <span className="conversational-correction-badge">已修订</span>}
+                </div>
                 <button
                   aria-label={`更正这条经历：${entry.summary}`}
                   disabled={controller.pending || !canAnswer}
                   type="button"
                   onClick={() => {
-                    controller.setDraft(`更正「${entry.summary}」（${entry.dateLabel}）：`);
+                    controller.beginEvidenceCorrection(entry.id);
                     focusComposer();
                   }}
                 >
diff --git a/frontend/src/hooks/use-conversational-rectification.ts b/frontend/src/hooks/use-conversational-rectification.ts
index cf137590..d10a2c8a 100644
--- a/frontend/src/hooks/use-conversational-rectification.ts
+++ b/frontend/src/hooks/use-conversational-rectification.ts
@@ -16,11 +16,13 @@ import type {
 type EvidenceDomain = NonNullable<
   ConversationalRectificationTurn["evidenceRequest"]
 >["domains"][number];
+type EvidenceRecapEntry = ConversationalRectificationTurn["evidenceRecap"][number];
 
 export type ConversationalRectificationControllerSnapshot = Readonly<{
   turn: ConversationalRectificationTurn | null;
   draft: string;
   selectedDomain: EvidenceDomain | null;
+  correctionTarget: EvidenceRecapEntry | null;
   pending: boolean;
   error: string;
 }>;
@@ -33,6 +35,8 @@ export type ConversationalRectificationController = ConversationalRectificationC
   synchronizeInitialTurn(turn: ConversationalRectificationTurn | null): void;
   setDraft(value: string): void;
   selectDomain(domain: EvidenceDomain | null): void;
+  beginEvidenceCorrection(evidenceId: string): void;
+  cancelEvidenceCorrection(): void;
   start(pendingConsultationQuestion?: string | null): MutationResult;
   resume(): MutationResult;
   answer(domain?: EvidenceDomain): MutationResult;
@@ -97,6 +101,7 @@ export function createConversationalRectificationController(
     turn: input.initialTurn ?? null,
     draft: "",
     selectedDomain: null,
+    correctionTarget: null,
     pending: false,
     error: "",
   };
@@ -123,10 +128,16 @@ export function createConversationalRectificationController(
       : snapshot.selectedDomain && turn.evidenceRequest?.domains.includes(snapshot.selectedDomain)
         ? snapshot.selectedDomain
         : null;
+    const correctionTarget = clearDraft
+      ? null
+      : snapshot.correctionTarget
+        ? turn.evidenceRecap.find((entry) => entry.id === snapshot.correctionTarget?.id) ?? null
+        : null;
     patch({
       turn,
       error: "",
       selectedDomain,
+      correctionTarget,
       ...(clearDraft ? { draft: "" } : {}),
     });
     try {
@@ -216,6 +227,7 @@ export function createConversationalRectificationController(
     get turn() { return snapshot.turn; },
     get draft() { return snapshot.draft; },
     get selectedDomain() { return snapshot.selectedDomain; },
+    get correctionTarget() { return snapshot.correctionTarget; },
     get pending() { return snapshot.pending; },
     get error() { return snapshot.error; },
     getSnapshot: () => snapshot,
@@ -229,13 +241,27 @@ export function createConversationalRectificationController(
         if (current === null) return;
         caseContext += 1;
         activeMutation = null;
-        patch({ turn: null, draft: "", selectedDomain: null, pending: false, error: "" });
+        patch({
+          turn: null,
+          draft: "",
+          selectedDomain: null,
+          correctionTarget: null,
+          pending: false,
+          error: "",
+        });
         return;
       }
       if (current === null || current.caseId !== turn.caseId) {
         caseContext += 1;
         activeMutation = null;
-        patch({ turn, draft: "", selectedDomain: null, pending: false, error: "" });
+        patch({
+          turn,
+          draft: "",
+          selectedDomain: null,
+          correctionTarget: null,
+          pending: false,
+          error: "",
+        });
         return;
       }
       if (turn.turnVersion <= current.turnVersion) return;
@@ -246,6 +272,9 @@ export function createConversationalRectificationController(
           && turn.evidenceRequest?.domains.includes(snapshot.selectedDomain)
           ? snapshot.selectedDomain
           : null,
+        correctionTarget: snapshot.correctionTarget
+          ? turn.evidenceRecap.find((entry) => entry.id === snapshot.correctionTarget?.id) ?? null
+          : null,
       });
     },
     setDraft(value: string) {
@@ -254,6 +283,19 @@ export function createConversationalRectificationController(
     selectDomain(domain: EvidenceDomain | null) {
       patch({ selectedDomain: domain });
     },
+    beginEvidenceCorrection(evidenceId: string) {
+      const target = snapshot.turn?.evidenceRecap.find((entry) => entry.id === evidenceId);
+      if (!target || !snapshot.turn?.actions.includes("answer")) return;
+      patch({
+        correctionTarget: target,
+        // Keep the old fact visible in the correction banner, but out of the new raw evidence.
+        // Otherwise its old date can make an appended replacement date look ambiguous.
+        draft: "",
+      });
+    },
+    cancelEvidenceCorrection() {
+      patch({ correctionTarget: null, draft: "" });
+    },
     start(pendingConsultationQuestion: string | null = null) {
       return run({
         identity: {
@@ -277,7 +319,12 @@ export function createConversationalRectificationController(
       const turn = currentTurn();
       const answer = snapshot.draft.trim();
       if (!turn || !answer || !turn.actions.includes("answer")) return Promise.resolve(turn);
-      const payload = { answer, ...(domain ? { domain } : {}) };
+      const correctsEvidenceId = snapshot.correctionTarget?.id;
+      const payload = {
+        answer,
+        ...(domain ? { domain } : {}),
+        ...(correctsEvidenceId ? { correctsEvidenceId } : {}),
+      };
       return currentMutation("answer", payload, (current, actionId) => ({
         type: "answer",
         caseId: current.caseId,
@@ -285,6 +332,7 @@ export function createConversationalRectificationController(
         turnVersion: current.turnVersion,
         answer,
         ...(domain ? { domain } : {}),
+        ...(correctsEvidenceId ? { correctsEvidenceId } : {}),
       }), true);
     },
     pause() {
diff --git a/frontend/src/lib/conversational-rectification/contracts.ts b/frontend/src/lib/conversational-rectification/contracts.ts
index 061627f3..e9d73bae 100644
--- a/frontend/src/lib/conversational-rectification/contracts.ts
+++ b/frontend/src/lib/conversational-rectification/contracts.ts
@@ -42,6 +42,7 @@ export const conversationalRectificationCommandSchema = z.discriminatedUnion("ty
     type: z.literal("answer"),
     domain: evidenceDomainSchema.optional(),
     answer: z.string().trim().min(1).max(4_000),
+    correctsEvidenceId: z.string().uuid().optional(),
   }).strict(),
   actionCommandSchema.extend({
     type: z.literal("pause"),
@@ -81,6 +82,8 @@ const evidenceRecapEntrySchema = boundedJson(z.object({
   id: z.string().uuid(),
   summary: boundedNonblankText(1_000),
   dateLabel: boundedNonblankText(80),
+  // Optional so turns written before correction lineage was introduced still resume safely.
+  isCorrection: z.boolean().optional(),
 }).strict(), 4_096);
 const evidenceRecapSchema = boundedJson(
   z.array(evidenceRecapEntrySchema).max(20),
diff --git a/frontend/src/lib/conversational-rectification/evidence-extractor.ts b/frontend/src/lib/conversational-rectification/evidence-extractor.ts
index a1c47f9d..93b58dad 100644
--- a/frontend/src/lib/conversational-rectification/evidence-extractor.ts
+++ b/frontend/src/lib/conversational-rectification/evidence-extractor.ts
@@ -10,13 +10,14 @@ export type ExtractedLifeEventEvidence = {
   readonly datePrecision: "day" | "month" | "year" | "unknown";
   readonly extractionStatus: "clear" | "needs_clarification" | "corrected";
   readonly scoreable: boolean;
+  readonly correctsEvidenceIds: readonly string[];
 };
 
 export type ExtractLifeEventEvidenceInput = {
   readonly rawText: string;
   readonly sourceTurnId: string;
   readonly asOfDate: string;
-  readonly correctionOfEvidenceIds?: readonly string[];
+  readonly correctsEvidenceId?: string;
 };
 
 type ParsedDate = {
@@ -139,7 +140,7 @@ export function extractLifeEventEvidence(
   if (!input.rawText.trim()) throw new TypeError("life-event raw text is required");
   if (!input.sourceTurnId.trim()) throw new TypeError("source turn id is required");
   if (!/^\d{4}-\d{2}-\d{2}$/.test(input.asOfDate)) throw new TypeError("asOfDate must be YYYY-MM-DD");
-  const corrections = [...(input.correctionOfEvidenceIds ?? [])];
+  const correctionTargets = input.correctsEvidenceId ? [input.correctsEvidenceId] : [];
   const events: ExtractedLifeEventEvidence[] = [];
 
   for (const fragments of splitSentences(input.rawText.normalize("NFKC"))) {
@@ -155,7 +156,7 @@ export function extractLifeEventEvidence(
       const complete = summary !== missingEventSummary && date !== null && !unresolvedRelativeTime;
       const extractionStatus = !complete
         ? "needs_clarification"
-        : corrections.length > 0 ? "corrected" : "clear";
+        : correctionTargets.length > 0 ? "corrected" : "clear";
       events.push({
         id: evidenceId(input, events.length, summary),
         rawText: input.rawText,
@@ -165,6 +166,7 @@ export function extractLifeEventEvidence(
         datePrecision: date?.precision ?? "unknown",
         extractionStatus,
         scoreable: complete && !dateIsFuture(date, input.asOfDate),
+        correctsEvidenceIds: correctionTargets,
       });
     }
   }
diff --git a/frontend/src/lib/conversational-rectification/orchestrator.ts b/frontend/src/lib/conversational-rectification/orchestrator.ts
index 9e9f6f9f..210550f7 100644
--- a/frontend/src/lib/conversational-rectification/orchestrator.ts
+++ b/frontend/src/lib/conversational-rectification/orchestrator.ts
@@ -134,7 +134,7 @@ type MutableCommand = Extract<ConversationalRectificationCommand, {
 function commandFingerprint(command: MutableCommand): string {
   const identity = command.type === "answer"
     ? [command.type, command.caseId, command.actionId, command.turnVersion,
-        command.domain ?? null, command.answer]
+        command.domain ?? null, command.answer, command.correctsEvidenceId ?? null]
     : command.type === "confirm"
       ? [command.type, command.caseId, command.actionId, command.turnVersion, command.time]
       : [command.type, command.caseId, command.actionId, command.turnVersion];
@@ -165,8 +165,21 @@ function latestReceipt(value: LoadedConversationalRectificationCase): Validation
   return parsed.data;
 }
 
+export function effectiveLifeEventEvidence<
+  Evidence extends Readonly<{
+    id: string;
+    correctsEvidenceIds?: readonly string[];
+  }>,
+>(evidence: ReadonlyArray<Evidence>): ReadonlyArray<Evidence> {
+  const correctedIds = new Set<string>();
+  for (const item of evidence) {
+    for (const correctedId of item.correctsEvidenceIds ?? []) correctedIds.add(correctedId);
+  }
+  return evidence.filter((item) => !correctedIds.has(item.id));
+}
+
 function evidenceRecap(evidence: ReadonlyArray<LifeEventEvidenceInput>) {
-  return evidence.slice(-20).map((item) => ({
+  return effectiveLifeEventEvidence(evidence).slice(-20).map((item) => ({
     id: item.id,
     summary: item.eventSummary,
     dateLabel: item.dateValue
@@ -174,6 +187,7 @@ function evidenceRecap(evidence: ReadonlyArray<LifeEventEvidenceInput>) {
         ? `${item.dateValue}（未来，仅作背景）`
         : item.dateValue
       : "日期待补充",
+    ...((item.correctsEvidenceIds?.length ?? 0) > 0 ? { isCorrection: true } : {}),
   }));
 }
 
@@ -238,6 +252,7 @@ function privateCandidateFromPacket(input: {
   readonly packet: RectificationTechnicalPacket;
   readonly resultId: string | null;
   readonly iteration: number;
+  readonly forceCollecting?: boolean;
 }): PrivateCandidate {
   const packet = input.packet;
   const parsed = privateCandidateSchema.safeParse({
@@ -257,7 +272,9 @@ function privateCandidateFromPacket(input: {
     suggestedDomains: packet.suggestedDomains.map((item) => item.domain),
     futureWindows: packet.futureWindows,
     workingState: {
-      phase: packet.candidate.status === "ready_for_confirmation" ? "ready" : "collecting_evidence",
+      phase: input.forceCollecting
+        ? "collecting_evidence"
+        : packet.candidate.status === "ready_for_confirmation" ? "ready" : "collecting_evidence",
       iteration: input.iteration,
       notes: [],
     },
@@ -320,18 +337,23 @@ function nonScoringTurn(input: {
   readonly domain?: RectificationEvidenceDomain;
   readonly directionChange: boolean;
   readonly scoringFallback?: boolean;
+  readonly correctionClarificationPacket?: RectificationTechnicalPacket;
 }): { readonly turn: ConversationalRectificationTurn; readonly receipt: ValidationReceipt } {
   const allEvidence = [...input.current.eventEvidence, ...input.newEvidence];
   const hasFuture = input.newEvidence.some((item) => item.extractionStatus !== "needs_clarification"
     && item.scoreable === false && item.dateValue !== null);
-  const narrative = input.scoringFallback
+  const narrative = input.correctionClarificationPacket
+    ? "这条更正已保存，原记录已经停止参与候选评分。更正后的事件时间还不够清楚，请补充大约年份、月份和发生了什么；在补清之前不会沿用旧证据推进确认。"
+    : input.scoringFallback
     ? "本轮原文已安全保存，但新的专业解释未通过事实一致性校验，因此候选没有推进。请稍后重试，或继续补充一件已经发生并带有年月的事件。"
     : input.directionChange
       ? "好的，我们不沿用不符合你的方向。你可以自由描述另一件已经发生的生活变化，尽量写明年月；我会根据事实继续，而不是让你选择宽泛年份。"
       : hasFuture
         ? "已保存这段描述。未来事件只能作为背景，不能用于校正评分；请再说一件已经发生的事件，并尽量写明年月。"
         : "我已保存你的原话，但还缺少可用于区分候选的明确时间。请用自己的话补充这件已经发生的事大约是哪一年、哪一月；不需要选择固定答案。";
-  const status = input.current.status === "confirming" ? "confirming" : "active";
+  const status = input.correctionClarificationPacket
+    ? "active" as const
+    : input.current.status === "confirming" ? "confirming" as const : "active" as const;
   const actions = actionsFor(status);
   const evidenceRequest = status === "confirming" && input.current.latestTurn.evidenceRequest === null
     ? null
@@ -348,6 +370,13 @@ function nonScoringTurn(input: {
     evidenceRequest,
     evidenceRecap: evidenceRecap(allEvidence),
     actions,
+    ...(input.correctionClarificationPacket ? {
+      candidate: {
+        ...projectRectificationTechnicalPacket(input.correctionClarificationPacket).candidate,
+        status: "pending_validation" as const,
+      },
+      technicalReceipt: exactTechnicalReceipt(input.correctionClarificationPacket),
+    } : {}),
   });
   if (!parsed.success) throw new ConversationalRectificationError("service_unavailable");
   return {
@@ -416,8 +445,10 @@ export function createConversationalRectificationService(
         rawText: command.answer,
         sourceTurnId: command.actionId,
         asOfDate: ports.asOfDate(),
+        correctsEvidenceId: command.correctsEvidenceId,
       }).map((item) => ({
         ...item,
+        correctsEvidenceIds: [...item.correctsEvidenceIds],
         domain: item.domain === "other" && command.domain && command.domain !== "other"
           ? command.domain
           : item.domain,
@@ -602,11 +633,62 @@ export function createConversationalRectificationService(
       }
       requireExactVersion(current, command.turnVersion);
 
+      if (command.correctsEvidenceId
+        && !effectiveLifeEventEvidence(current.eventEvidence)
+          .some((item) => item.id === command.correctsEvidenceId)) {
+        throw new ConversationalRectificationError("action_conflict");
+      }
+
       const scoreableEvidence = evidence.filter((item) => item.scoreable === true
         && item.extractionStatus !== "needs_clarification");
       const explicitDirectionChange = explicitDirectionChangePattern.test(command.answer);
       const directionChange = explicitDirectionChange
         || (scoreableEvidence.length === 0 && genericUncertaintyPattern.test(command.answer));
+      const unclearCorrection = Boolean(command.correctsEvidenceId)
+        && evidence.some((item) => item.extractionStatus === "needs_clarification");
+      if (unclearCorrection) {
+        try {
+          const allScoreable = effectiveLifeEventEvidence([...current.eventEvidence, ...evidence])
+            .filter((item) => item.scoreable === true
+              && item.extractionStatus !== "needs_clarification"
+              && !evidencePredatesBirthDate(item, current.declaredBirthInput.birthDate));
+          const computed = await ports.buildTechnicalPacket({
+            userId,
+            caseId: command.caseId,
+            asOfDate: ports.asOfDate(),
+            declaredBirthInput: current.declaredBirthInput,
+            privateCandidate: current.privateCandidate,
+            evidence: allScoreable,
+          });
+          const next = nonScoringTurn({
+            current,
+            newEvidence: evidence,
+            domain: command.domain,
+            directionChange: false,
+            correctionClarificationPacket: computed.packet,
+          });
+          const privateCandidate = privateCandidateFromPacket({
+            packet: computed.packet,
+            resultId: null,
+            iteration: (current.privateCandidate.workingState?.iteration ?? 0) + 1,
+            forceCollecting: true,
+          });
+          const saved = await ports.store.saveTurn({
+            userId,
+            caseId: command.caseId,
+            expectedVersion: command.turnVersion,
+            actionId: command.actionId,
+            commandFingerprint: fingerprint,
+            turn: next.turn,
+            evidence,
+            validationReceipt: next.receipt,
+            privateCandidate,
+          });
+          return publicTurn(saved);
+        } catch (error) {
+          throw safeFailure(error);
+        }
+      }
       if (directionChange || scoreableEvidence.length === 0) {
         const next = nonScoringTurn({
           current,
@@ -633,7 +715,7 @@ export function createConversationalRectificationService(
       }
 
       try {
-        const allScoreable = [...current.eventEvidence, ...evidence]
+        const allScoreable = effectiveLifeEventEvidence([...current.eventEvidence, ...evidence])
           .filter((item) => item.scoreable === true
             && item.extractionStatus !== "needs_clarification"
             && !evidencePredatesBirthDate(item, current.declaredBirthInput.birthDate));
diff --git a/frontend/src/lib/conversational-rectification/persistence-contracts.ts b/frontend/src/lib/conversational-rectification/persistence-contracts.ts
index 38577637..db4ab0f4 100644
--- a/frontend/src/lib/conversational-rectification/persistence-contracts.ts
+++ b/frontend/src/lib/conversational-rectification/persistence-contracts.ts
@@ -181,6 +181,8 @@ export const validationReceiptSchema = boundedJson(z.object({
 }).strict(), 8_192);
 export type ValidationReceipt = z.infer<typeof validationReceiptSchema>;
 
+const correctionEvidenceIdsSchema = boundedJson(z.array(uuidSchema).max(1), 64);
+
 export const lifeEventEvidenceSchema = boundedJson(z.object({
   id: uuidSchema,
   rawText: boundedText(4_000),
@@ -190,7 +192,17 @@ export const lifeEventEvidenceSchema = boundedJson(z.object({
   datePrecision: z.enum(["day", "month", "year", "range", "unknown"]),
   extractionStatus: z.enum(["clear", "needs_clarification", "corrected"]),
   scoreable: z.boolean().optional(),
-}).strict(), 16_384);
+  // Optional only for rows written before durable correction lineage existed.
+  correctsEvidenceIds: correctionEvidenceIdsSchema.optional(),
+}).strict().superRefine((value, context) => {
+  const correctionCount = value.correctsEvidenceIds?.length ?? 0;
+  if (value.extractionStatus === "corrected" && correctionCount !== 1) {
+    context.addIssue({ code: "custom", message: "corrected evidence requires one target" });
+  }
+  if (value.extractionStatus === "clear" && correctionCount !== 0) {
+    context.addIssue({ code: "custom", message: "clear evidence cannot correct another row" });
+  }
+}), 16_384);
 export type LifeEventEvidence = z.infer<typeof lifeEventEvidenceSchema>;
 
 const mutationKindSchema = z.enum([
diff --git a/frontend/supabase/migrations/20260720010000_conversational_rectification_schema.sql b/frontend/supabase/migrations/20260720010000_conversational_rectification_schema.sql
index 2f20ef68..af9b3399 100644
--- a/frontend/supabase/migrations/20260720010000_conversational_rectification_schema.sql
+++ b/frontend/supabase/migrations/20260720010000_conversational_rectification_schema.sql
@@ -293,7 +293,7 @@ as $$
       where pg_catalog.jsonb_typeof(item) <> 'object'
         or pg_catalog.octet_length(item::text) > 4096
         or not public.conversational_rectification_has_only_keys(
-          item, array['id', 'summary', 'dateLabel']::text[]
+          item, array['id', 'summary', 'dateLabel', 'isCorrection']::text[]
         )
         or not (item ?& array['id', 'summary', 'dateLabel']::text[])
         or pg_catalog.jsonb_typeof(item -> 'id') <> 'string'
@@ -312,6 +312,10 @@ as $$
         or public.conversational_rectification_text_is_nonblank(
           item ->> 'dateLabel'
         ) is not true
+        or (
+          item ? 'isCorrection'
+          and pg_catalog.jsonb_typeof(item -> 'isCorrection') <> 'boolean'
+        )
     );
 $$;
 
@@ -423,7 +427,7 @@ begin
       p_value,
       array[
         'id', 'rawText', 'domain', 'eventSummary', 'dateValue',
-        'datePrecision', 'extractionStatus', 'scoreable'
+        'datePrecision', 'extractionStatus', 'scoreable', 'correctsEvidenceIds'
       ]::text[]
     )
     or not (p_value ?& array[
@@ -473,6 +477,32 @@ begin
     or (
       p_value ? 'scoreable'
       and pg_catalog.jsonb_typeof(p_value -> 'scoreable') is distinct from 'boolean'
+    )
+    or (
+      p_value ? 'correctsEvidenceIds'
+      and (
+        pg_catalog.jsonb_typeof(p_value -> 'correctsEvidenceIds') is distinct from 'array'
+        or pg_catalog.octet_length((p_value -> 'correctsEvidenceIds')::text) > 64
+        or pg_catalog.jsonb_array_length(p_value -> 'correctsEvidenceIds') > 1
+        or exists (
+          select 1
+          from pg_catalog.jsonb_array_elements(p_value -> 'correctsEvidenceIds') target
+          where pg_catalog.jsonb_typeof(target) is distinct from 'string'
+            or public.conversational_rectification_valid_uuid_text(target #>> '{}') is not true
+        )
+      )
+    )
+    or (
+      p_value ->> 'extractionStatus' = 'corrected'
+      and (
+        not (p_value ? 'correctsEvidenceIds')
+        or pg_catalog.jsonb_array_length(p_value -> 'correctsEvidenceIds') <> 1
+      )
+    )
+    or (
+      p_value ->> 'extractionStatus' = 'clear'
+      and p_value ? 'correctsEvidenceIds'
+      and pg_catalog.jsonb_array_length(p_value -> 'correctsEvidenceIds') <> 0
     ) then
     return false;
   end if;
@@ -1209,6 +1239,14 @@ create table if not exists public.birth_time_rectification_event_evidence (
   extraction_status text not null check (
     extraction_status in ('clear', 'needs_clarification', 'corrected')
   ),
+  corrects_evidence_ids uuid[] not null default '{}'::uuid[] check (
+    pg_catalog.cardinality(corrects_evidence_ids) <= 1
+    and (
+      (extraction_status = 'corrected' and pg_catalog.cardinality(corrects_evidence_ids) = 1)
+      or (extraction_status = 'clear' and pg_catalog.cardinality(corrects_evidence_ids) = 0)
+      or extraction_status = 'needs_clarification'
+    )
+  ),
   scoreable boolean not null default false,
   created_at timestamptz not null default now(),
   foreign key (case_id, source_turn_id)
diff --git a/frontend/supabase/migrations/20260720030000_conversational_rectification_transitions.sql b/frontend/supabase/migrations/20260720030000_conversational_rectification_transitions.sql
index b91999b3..ccdc79f8 100644
--- a/frontend/supabase/migrations/20260720030000_conversational_rectification_transitions.sql
+++ b/frontend/supabase/migrations/20260720030000_conversational_rectification_transitions.sql
@@ -152,7 +152,8 @@ begin
           'dateValue', evidence.date_value,
           'datePrecision', evidence.date_precision,
           'extractionStatus', evidence.extraction_status,
-          'scoreable', evidence.scoreable
+          'scoreable', evidence.scoreable,
+          'correctsEvidenceIds', pg_catalog.to_jsonb(evidence.corrects_evidence_ids)
         ) order by evidence.created_at, evidence.id
       )
       from public.birth_time_rectification_event_evidence evidence
@@ -273,7 +274,8 @@ begin
       'dateValue', evidence.date_value,
       'datePrecision', evidence.date_precision,
       'extractionStatus', evidence.extraction_status,
-      'scoreable', evidence.scoreable
+      'scoreable', evidence.scoreable,
+      'correctsEvidenceIds', pg_catalog.to_jsonb(evidence.corrects_evidence_ids)
     ) order by evidence.created_at, evidence.id
   ), '[]'::jsonb) into v_existing_evidence
   from public.birth_time_rectification_event_evidence evidence
@@ -294,7 +296,10 @@ begin
       'scoreable', case
         when item.value ? 'scoreable' then (item.value ->> 'scoreable')::boolean
         else false
-      end
+      end,
+      'correctsEvidenceIds', coalesce(
+        item.value -> 'correctsEvidenceIds', '[]'::jsonb
+      )
     ) order by item.ordinality
   ), '[]'::jsonb) into v_new_evidence
   from pg_catalog.jsonb_array_elements(p_new_evidence) with ordinality item(value, ordinality);
@@ -634,6 +639,32 @@ begin
     or pg_catalog.jsonb_path_exists(p_turn, '$.**.systemPrompt') then
     raise exception 'conversational_action_conflict' using errcode = 'P0001';
   end if;
+  -- Correction targets are account-case evidence tips. Under the locked case version,
+  -- reject missing, cross-case, or already-retired targets before writing a turn.
+  if exists (
+    select 1
+    from pg_catalog.jsonb_array_elements(p_evidence) item(value)
+    cross join lateral pg_catalog.jsonb_array_elements_text(
+      coalesce(item.value -> 'correctsEvidenceIds', '[]'::jsonb)
+    ) correction(value)
+    where not exists (
+      select 1
+      from public.birth_time_rectification_event_evidence target
+      where target.case_id = p_case_id
+        and target.id = correction.value::uuid
+    )
+    or exists (
+      select 1
+      from public.birth_time_rectification_event_evidence evidence
+      cross join lateral pg_catalog.unnest(
+        evidence.corrects_evidence_ids
+      ) retired(target_id)
+      where evidence.case_id = p_case_id
+        and retired.target_id = correction.value::uuid
+    )
+  ) then
+    raise exception 'conversational_action_conflict' using errcode = 'P0001';
+  end if;
   if public.conversational_rectification_case_fits_load_limits(
     p_user_id, p_case_id, p_turn ->> 'status', p_expected_version + 1,
     p_turn, p_private_candidate, p_evidence, p_validation_receipt
@@ -655,7 +686,7 @@ begin
 
   insert into public.birth_time_rectification_event_evidence (
     id, case_id, source_turn_id, raw_text, domain, event_summary,
-    date_value, date_precision, extraction_status, scoreable
+    date_value, date_precision, extraction_status, corrects_evidence_ids, scoreable
   )
   select
     (item ->> 'id')::uuid, p_case_id, v_turn_id, item ->> 'rawText',
@@ -664,6 +695,12 @@ begin
       else item ->> 'dateValue' end,
     item ->> 'datePrecision',
     item ->> 'extractionStatus',
+    case when item ? 'correctsEvidenceIds' then array(
+      select correction.value::uuid
+      from pg_catalog.jsonb_array_elements_text(
+        item -> 'correctsEvidenceIds'
+      ) correction(value)
+    ) else '{}'::uuid[] end,
     case when item ? 'scoreable' then (item ->> 'scoreable')::boolean
       else false end
   from pg_catalog.jsonb_array_elements(p_evidence) item;
diff --git a/frontend/tests/conversational-evidence-extractor.test.ts b/frontend/tests/conversational-evidence-extractor.test.ts
index 720c0779..c78573c4 100644
--- a/frontend/tests/conversational-evidence-extractor.test.ts
+++ b/frontend/tests/conversational-evidence-extractor.test.ts
@@ -112,14 +112,32 @@ test("marks a future event as context-only and non-scoreable", () => {
 });
 
 test("clear replacement evidence is explicitly marked corrected", () => {
+  const targetId = "00000000-0000-4000-8000-000000000611";
   const [evidence] = extractLifeEventEvidence({
     rawText: "更正：2020年11月离职",
     sourceTurnId,
     asOfDate: "2026-07-20",
-    correctionOfEvidenceIds: ["00000000-0000-4000-8000-000000000611"],
+    correctsEvidenceId: targetId,
   });
 
   assert.equal(evidence?.eventSummary, "离职");
   assert.equal(evidence?.extractionStatus, "corrected");
   assert.equal(evidence?.scoreable, true);
+  assert.deepEqual(evidence?.correctsEvidenceIds, [targetId]);
+  assert.equal(lifeEventEvidenceSchema.safeParse(evidence).success, true);
+});
+
+test("an unclear correction keeps durable lineage while withholding both old and new scoring", () => {
+  const targetId = "00000000-0000-4000-8000-000000000611";
+  const [evidence] = extractLifeEventEvidence({
+    rawText: "更正：具体年月记不清",
+    sourceTurnId,
+    asOfDate: "2026-07-20",
+    correctsEvidenceId: targetId,
+  });
+
+  assert.equal(evidence?.extractionStatus, "needs_clarification");
+  assert.equal(evidence?.scoreable, false);
+  assert.deepEqual(evidence?.correctsEvidenceIds, [targetId]);
+  assert.equal(lifeEventEvidenceSchema.safeParse(evidence).success, true);
 });
diff --git a/frontend/tests/conversational-rectification-component.test.ts b/frontend/tests/conversational-rectification-component.test.ts
index 2fa729c4..51528a9f 100644
--- a/frontend/tests/conversational-rectification-component.test.ts
+++ b/frontend/tests/conversational-rectification-component.test.ts
@@ -51,6 +51,7 @@ const turn: ConversationalRectificationTurn = {
     id: "00000000-0000-4000-8000-000000000822",
     summary: "开始第一份长期工作",
     dateLabel: "2021 年 7 月",
+    isCorrection: false,
   }],
   actions: ["answer", "pause", "abandon", "confirm"],
   pendingConsultationQuestion: "我适合什么时候换工作？",
@@ -61,13 +62,18 @@ function controller(overrides: Partial<ConversationalRectificationController> =
     turn,
     draft: "",
     selectedDomain: null,
+    correctionTarget: null,
     pending: false,
     error: "",
-    getSnapshot: () => ({ turn, draft: "", selectedDomain: null, pending: false, error: "" }),
+    getSnapshot: () => ({
+      turn, draft: "", selectedDomain: null, correctionTarget: null, pending: false, error: "",
+    }),
     subscribe: () => () => undefined,
     synchronizeInitialTurn: () => undefined,
     setDraft: () => undefined,
     selectDomain: () => undefined,
+    beginEvidenceCorrection: () => undefined,
+    cancelEvidenceCorrection: () => undefined,
     start: async () => turn,
     resume: async () => turn,
     answer: async () => turn,
@@ -112,6 +118,35 @@ test("evidence is correctable, technical receipts stay visible, and confirmation
   assert.match(markup, /放弃本次校正/);
 });
 
+test("correction mode identifies its durable target, can be cancelled, and marks revised recaps", () => {
+  const revisedTurn = {
+    ...turn,
+    evidenceRecap: [{ ...turn.evidenceRecap[0]!, isCorrection: true }],
+  } satisfies ConversationalRectificationTurn;
+  const correctionTarget = revisedTurn.evidenceRecap[0]!;
+  const markup = renderToStaticMarkup(React.createElement(
+    ConversationalRectificationSurface,
+    { controller: controller({
+      turn: revisedTurn,
+      draft: "更正：其实是 2020 年 11 月离职",
+      correctionTarget,
+      getSnapshot: () => ({
+        turn: revisedTurn,
+        draft: "更正：其实是 2020 年 11 月离职",
+        selectedDomain: null,
+        correctionTarget,
+        pending: false,
+        error: "",
+      }),
+    }) },
+  ));
+
+  assert.match(markup, /正在更正/);
+  assert.match(markup, /开始第一份长期工作/);
+  assert.match(markup, /取消更正/);
+  assert.match(markup, /已修订/);
+});
+
 test("pending markup and responsive CSS expose accessibility contracts", () => {
   const pendingController = controller({
     pending: true,
@@ -120,6 +155,7 @@ test("pending markup and responsive CSS expose accessibility contracts", () => {
       turn,
       draft: "保留中的文字",
       selectedDomain: "career",
+      correctionTarget: null,
       pending: true,
       error: "",
     }),
@@ -559,7 +595,12 @@ test("real Chromium at 390px verifies layout, keyboard focus, pause affordance,
         datePrecision: "month_preferred",
         freeTextAllowed: true,
       },
-      evidenceRecap: [],
+      evidenceRecap: status === "abandoned" ? [] : [{
+        id: "00000000-0000-4000-8000-000000000822",
+        summary: "开始第一份长期工作",
+        dateLabel: "2021-07",
+        isCorrection: false,
+      }],
       actions: status === "abandoned" ? [] : status === "paused"
         ? ["answer", "abandon"]
         : ["answer", "pause", "abandon"],
@@ -662,6 +703,22 @@ test("real Chromium at 390px verifies layout, keyboard focus, pause affordance,
     assert.ok(layout.surfaceWidth <= 366, `surface overflowed padded viewport: ${layout.surfaceWidth}px`);
     assert.ok(layout.shortestButton >= 44, `shortest button was ${layout.shortestButton}px`);
 
+    await cdp.evaluate("document.querySelector('[aria-label^=\"更正这条经历\"]').click()");
+    await waitFor(
+      () => cdp?.evaluate<boolean>(`(() => {
+        const textarea = document.getElementById('conversational-rectification-answer');
+        return document.body.textContent.includes('正在更正')
+          && document.body.textContent.includes('开始第一份长期工作')
+          && textarea?.value === '';
+      })()`) ?? Promise.resolve(false),
+      "durable correction target selection without polluting the new answer",
+    );
+    await cdp.evaluate("[...document.querySelectorAll('button')].find((button) => button.textContent.includes('取消更正')).click()");
+    await waitFor(
+      () => cdp?.evaluate<boolean>("!document.body.textContent.includes('正在更正') && document.getElementById('conversational-rectification-answer').value === ''") ?? Promise.resolve(false),
+      "correction cancellation",
+    );
+
     await cdp.evaluate("document.querySelector('[data-evidence-domain=career]').click()");
     await waitFor(
       () => cdp?.evaluate<boolean>("document.activeElement?.id === 'conversational-rectification-answer'") ?? Promise.resolve(false),
diff --git a/frontend/tests/conversational-rectification-contracts.test.ts b/frontend/tests/conversational-rectification-contracts.test.ts
index 5c3fcced..8adc08ae 100644
--- a/frontend/tests/conversational-rectification-contracts.test.ts
+++ b/frontend/tests/conversational-rectification-contracts.test.ts
@@ -11,6 +11,7 @@ import {
 
 const actionId = "a9890e09-d535-46f0-9a36-86017515a5a1";
 const caseId = "77b29d28-c576-429e-9e3d-d0a90348e3cb";
+const evidenceId = "37e0e35e-cfdc-4c7a-8375-84310ee6bd42";
 
 function parseCommand(value: unknown) {
   return conversationalRectificationCommandSchema.safeParse(value).success;
@@ -46,6 +47,33 @@ test("bounds free-text answers and requires a strict HH:mm confirmation time", (
   assert.equal(parseCommand({ type: "confirm", caseId, actionId, turnVersion: 1, time: "24:00" }), false);
 });
 
+test("accepts one explicit UUID correction target and rejects ambiguous target shapes", () => {
+  assert.equal(parseCommand({
+    type: "answer",
+    caseId,
+    actionId,
+    turnVersion: 1,
+    answer: "更正：2020 年 11 月离职",
+    correctsEvidenceId: evidenceId,
+  }), true);
+  assert.equal(parseCommand({
+    type: "answer",
+    caseId,
+    actionId,
+    turnVersion: 1,
+    answer: "更正：2020 年 11 月离职",
+    correctsEvidenceId: "not-a-uuid",
+  }), false);
+  assert.equal(parseCommand({
+    type: "answer",
+    caseId,
+    actionId,
+    turnVersion: 1,
+    answer: "更正两条",
+    correctsEvidenceIds: [evidenceId, caseId],
+  }), false);
+});
+
 test("rejects client candidate scores and technical receipts", () => {
   assert.equal(parseCommand({
     type: "answer", caseId, actionId, turnVersion: 1, answer: "2019 年 7 月换了工作",
@@ -82,9 +110,10 @@ test("accepts only the exact public turn shape", () => {
       freeTextAllowed: true,
     },
     evidenceRecap: [{
-      id: "37e0e35e-cfdc-4c7a-8375-84310ee6bd42",
+      id: evidenceId,
       summary: "2019 年换工作",
       dateLabel: "2019-07",
+      isCorrection: true,
     }],
     actions: ["answer", "pause", "abandon"],
     pendingConsultationQuestion: null,
@@ -94,6 +123,10 @@ test("accepts only the exact public turn shape", () => {
   assert.equal(conversationalRectificationTurnSchema.safeParse({ ...turn, candidateScores: [0.99] }).success, false);
   assert.equal(conversationalRectificationTurnSchema.safeParse({ ...turn, candidate: { ...turn.candidate, score: 0.99 } }).success, false);
   assert.equal(conversationalRectificationTurnSchema.safeParse({ ...turn, technicalReceipt: { ...turn.technicalReceipt, rawModelOutput: "secret" } }).success, false);
+  assert.equal(conversationalRectificationTurnSchema.safeParse({
+    ...turn,
+    evidenceRecap: [{ ...turn.evidenceRecap[0], isCorrection: "yes" }],
+  }).success, false);
 });
 
 function assertNoReachableText(value: unknown, forbidden: string, seen = new Set<unknown>()) {
diff --git a/frontend/tests/conversational-rectification-controller.test.ts b/frontend/tests/conversational-rectification-controller.test.ts
index f0ed9611..78fc8ceb 100644
--- a/frontend/tests/conversational-rectification-controller.test.ts
+++ b/frontend/tests/conversational-rectification-controller.test.ts
@@ -14,6 +14,7 @@ import type {
 
 const caseId = "00000000-0000-4000-8000-000000000811";
 const otherCaseId = "00000000-0000-4000-8000-000000000819";
+const evidenceId = "00000000-0000-4000-8000-000000000818";
 const actionIds = [
   "00000000-0000-4000-8000-000000000812",
   "00000000-0000-4000-8000-000000000813",
@@ -51,6 +52,18 @@ function activeTurn(turnVersion = 2): ConversationalRectificationTurn {
   };
 }
 
+function correctableTurn(turnVersion = 2): ConversationalRectificationTurn {
+  return {
+    ...activeTurn(turnVersion),
+    evidenceRecap: [{
+      id: evidenceId,
+      summary: "开始第一份长期工作",
+      dateLabel: "2021-07",
+      isCorrection: false,
+    }],
+  };
+}
+
 function idFactory() {
   const ids = [...actionIds];
   return () => ids.shift() ?? assert.fail("unexpected action id allocation");
@@ -124,6 +137,70 @@ test("controller preserves the exact draft and stable action id across failures"
   assert.equal(commands[0]?.actionId, commands[1]?.actionId);
 });
 
+test("a correction target is explicit in the command identity, survives failure, and clears on success", async () => {
+  const commands: ConversationalRectificationCommand[] = [];
+  let fail = true;
+  const controller = createConversationalRectificationController({
+    initialTurn: correctableTurn(),
+    createActionId: idFactory(),
+    send: async (command) => {
+      commands.push(command);
+      if (fail) throw new Error("offline");
+      return { ...correctableTurn(3), evidenceRecap: [] };
+    },
+  });
+
+  controller.beginEvidenceCorrection(evidenceId);
+  assert.equal(controller.getSnapshot().correctionTarget?.id, evidenceId);
+  assert.equal(
+    controller.getSnapshot().draft,
+    "",
+    "the old summary/date must stay outside the submitted correction text",
+  );
+  controller.setDraft("更正：其实是 2020 年 11 月离职");
+  await assert.rejects(controller.answer());
+  assert.equal(controller.getSnapshot().correctionTarget?.id, evidenceId);
+  assert.equal(commands[0]?.type, "answer");
+  assert.equal(commands[0]?.type === "answer" ? commands[0].correctsEvidenceId : null, evidenceId);
+
+  fail = false;
+  await controller.answer();
+  assert.equal(commands[0]?.actionId, commands[1]?.actionId);
+  assert.equal(controller.getSnapshot().correctionTarget, null);
+  assert.equal(controller.getSnapshot().draft, "");
+});
+
+test("canceling correction removes the target and changing targets changes the replay identity", async () => {
+  const secondEvidenceId = "00000000-0000-4000-8000-000000000817";
+  const commands: ConversationalRectificationCommand[] = [];
+  const initial = {
+    ...correctableTurn(),
+    evidenceRecap: [
+      ...correctableTurn().evidenceRecap,
+      { id: secondEvidenceId, summary: "搬到另一座城市", dateLabel: "2022-03" },
+    ],
+  } satisfies ConversationalRectificationTurn;
+  const controller = createConversationalRectificationController({
+    initialTurn: initial,
+    createActionId: idFactory(),
+    send: async (command) => {
+      commands.push(command);
+      throw new Error("offline");
+    },
+  });
+
+  controller.beginEvidenceCorrection(evidenceId);
+  controller.setDraft("更正：其实是 2020 年 11 月离职");
+  await assert.rejects(controller.answer());
+  controller.beginEvidenceCorrection(secondEvidenceId);
+  controller.setDraft("更正：其实是 2022 年 3 月搬家");
+  await assert.rejects(controller.answer());
+  assert.notEqual(commands[0]?.actionId, commands[1]?.actionId);
+  controller.cancelEvidenceCorrection();
+  assert.equal(controller.getSnapshot().correctionTarget, null);
+  assert.equal(controller.getSnapshot().draft, "");
+});
+
 test("a stale answer resumes the latest turn and returns recovered state without clearing text", async () => {
   const commands: ConversationalRectificationCommand[] = [];
   const recovered = activeTurn(5);
@@ -173,6 +250,32 @@ test("a stale recovery retains the selected domain only while the recovered turn
   assert.equal(controller.getSnapshot().selectedDomain, "career");
 });
 
+test("stale recovery retains a correction only when the latest effective recap still contains it", async () => {
+  const recoveryCases: Array<[ConversationalRectificationTurn, string | null]> = [
+    [correctableTurn(5), evidenceId],
+    [{ ...correctableTurn(5), evidenceRecap: [] }, null],
+  ];
+  for (const [recovered, expected] of recoveryCases) {
+    const controller = createConversationalRectificationController({
+      initialTurn: correctableTurn(2),
+      createActionId: idFactory(),
+      send: async (command) => {
+        if (command.type === "answer") {
+          throw new ConversationalRectificationRequestError(409, "stale_turn", "请加载最新进度后再试。");
+        }
+        return recovered;
+      },
+    });
+    controller.beginEvidenceCorrection(evidenceId);
+    controller.setDraft("更正：其实是 2020 年 11 月离职");
+
+    await controller.answer();
+
+    assert.equal(controller.getSnapshot().correctionTarget?.id ?? null, expected);
+    assert.equal(controller.getSnapshot().draft, "更正：其实是 2020 年 11 月离职");
+  }
+});
+
 test("a changed payload receives a different action id after a failed send", async () => {
   const commands: ConversationalRectificationCommand[] = [];
   const controller = createConversationalRectificationController({
@@ -289,6 +392,7 @@ test("a case switch detaches the old mutation so the new case can mutate indepen
     turn: caseBTurn,
     draft: "",
     selectedDomain: null,
+    correctionTarget: null,
     pending: false,
     error: "",
   });
@@ -332,6 +436,7 @@ test("synchronizing to no case detaches an ordinary failure without publishing i
     turn: null,
     draft: "",
     selectedDomain: null,
+    correctionTarget: null,
     pending: false,
     error: "",
   });
@@ -342,6 +447,7 @@ test("synchronizing to no case detaches an ordinary failure without publishing i
     turn: null,
     draft: "",
     selectedDomain: null,
+    correctionTarget: null,
     pending: false,
     error: "",
   });
diff --git a/frontend/tests/conversational-rectification-orchestrator.test.ts b/frontend/tests/conversational-rectification-orchestrator.test.ts
index 491377bb..b1e47e03 100644
--- a/frontend/tests/conversational-rectification-orchestrator.test.ts
+++ b/frontend/tests/conversational-rectification-orchestrator.test.ts
@@ -487,6 +487,137 @@ test("clear historical evidence is extracted, scored, narrated, recapped, and at
   assert.ok(value.events.includes("score-packet"));
 });
 
+test("evidence corrections are append-only while recap and scoring use only the effective lineage tip", async () => {
+  const value = harness({ readyAfterEvidenceCount: 99 });
+  await start(value, null);
+  await value.service.answer(userId, {
+    type: "answer",
+    caseId: startActionId,
+    actionId: answerActionId,
+    turnVersion: 0,
+    answer: "2019年7月开始第一份工作",
+  });
+  const firstId = value.cases.get(startActionId)?.row.eventEvidence[0]?.id;
+  assert.ok(firstId);
+
+  const corrected = await value.service.answer(userId, {
+    type: "answer",
+    caseId: startActionId,
+    actionId: secondAnswerActionId,
+    turnVersion: 1,
+    answer: "更正：其实是2020年11月离职",
+    correctsEvidenceId: firstId,
+  });
+  const secondId = value.cases.get(startActionId)?.row.eventEvidence[1]?.id;
+  assert.ok(secondId);
+  assert.deepEqual(corrected.evidenceRecap, [{
+    id: secondId,
+    summary: "其实是离职",
+    dateLabel: "2020-11",
+    isCorrection: true,
+  }]);
+
+  const twiceCorrected = await value.service.answer(userId, {
+    type: "answer",
+    caseId: startActionId,
+    actionId: thirdAnswerActionId,
+    turnVersion: 2,
+    answer: "更正：准确的是2021年2月入职",
+    correctsEvidenceId: secondId,
+  });
+
+  const stored = value.cases.get(startActionId)?.row.eventEvidence ?? [];
+  assert.equal(stored.length, 3, "the audit lineage must remain append-only");
+  assert.deepEqual(stored[0]?.correctsEvidenceIds ?? [], []);
+  assert.deepEqual(stored[1]?.correctsEvidenceIds, [firstId]);
+  assert.deepEqual(stored[2]?.correctsEvidenceIds, [secondId]);
+  assert.deepEqual(twiceCorrected.evidenceRecap, [{
+    id: stored[2]?.id,
+    summary: "准确的是入职",
+    dateLabel: "2021-02",
+    isCorrection: true,
+  }]);
+  assert.deepEqual(value.packetEvidenceCounts, [0, 1, 1, 1]);
+});
+
+test("an unclear correction immediately retires the wrong fact and stays retired after later turns", async () => {
+  const value = harness();
+  await start(value, null);
+  await value.service.answer(userId, {
+    type: "answer", caseId: startActionId, actionId: answerActionId,
+    turnVersion: 0, answer: "2019年7月开始第一份工作",
+  });
+  const wrongId = value.cases.get(startActionId)?.row.eventEvidence[0]?.id;
+  assert.ok(wrongId);
+
+  const clarification = await value.service.answer(userId, {
+    type: "answer", caseId: startActionId, actionId: secondAnswerActionId,
+    turnVersion: 1, answer: "更正：具体年月记不清", correctsEvidenceId: wrongId,
+  });
+  const unclear = value.cases.get(startActionId)?.row.eventEvidence[1];
+  assert.equal(unclear?.extractionStatus, "needs_clarification");
+  assert.equal(unclear?.scoreable, false);
+  assert.deepEqual(unclear?.correctsEvidenceIds, [wrongId]);
+  assert.deepEqual(clarification.evidenceRecap, [{
+    id: unclear?.id,
+    summary: "具体年月记不清",
+    dateLabel: "日期待补充",
+    isCorrection: true,
+  }]);
+  assert.equal(clarification.status, "active");
+  assert.equal(clarification.candidate.status, "pending_validation");
+  assert.equal(clarification.actions.includes("confirm"), false);
+  assert.equal(value.cases.get(startActionId)?.row.privateCandidate.resultId, null);
+  assert.deepEqual(
+    value.cases.get(startActionId)?.row.privateCandidate.scoredHistoricalEvidence ?? [],
+    [],
+  );
+  assert.deepEqual(value.packetEvidenceCounts, [0, 1, 0]);
+
+  const later = await value.service.answer(userId, {
+    type: "answer", caseId: startActionId, actionId: thirdAnswerActionId,
+    turnVersion: 2, answer: "2022年3月搬家",
+  });
+  assert.deepEqual(value.packetEvidenceCounts, [0, 1, 0, 1]);
+  assert.equal(later.evidenceRecap.some((item) => item.id === wrongId), false);
+  assert.equal(later.evidenceRecap.some((item) => item.id === unclear?.id), true);
+});
+
+test("missing or already-retired correction targets fail closed without advancing or persisting", async () => {
+  const value = harness({ readyAfterEvidenceCount: 99 });
+  await start(value, null);
+  const before = [...value.mutations];
+  await assert.rejects(value.service.answer(userId, {
+    type: "answer", caseId: startActionId, actionId: answerActionId,
+    turnVersion: 0, answer: "更正：2020年11月离职",
+    correctsEvidenceId: "00000000-0000-4000-8000-000000000799",
+  }), (error: unknown) => error instanceof ConversationalRectificationError
+    && error.code === "action_conflict");
+  assert.deepEqual(value.mutations, before);
+  assert.equal(value.cases.get(startActionId)?.row.turnVersion, 0);
+  assert.equal(value.cases.get(startActionId)?.row.eventEvidence.length, 0);
+
+  await value.service.answer(userId, {
+    type: "answer", caseId: startActionId, actionId: secondAnswerActionId,
+    turnVersion: 0, answer: "2019年7月开始第一份工作",
+  });
+  const firstId = value.cases.get(startActionId)?.row.eventEvidence[0]?.id;
+  assert.ok(firstId);
+  await value.service.answer(userId, {
+    type: "answer", caseId: startActionId, actionId: thirdAnswerActionId,
+    turnVersion: 1, answer: "更正：2020年11月离职", correctsEvidenceId: firstId,
+  });
+  const versionBeforeConflict = value.cases.get(startActionId)?.row.turnVersion;
+  const evidenceBeforeConflict = value.cases.get(startActionId)?.row.eventEvidence.length;
+  await assert.rejects(value.service.answer(userId, {
+    type: "answer", caseId: startActionId, actionId: fourthAnswerActionId,
+    turnVersion: 2, answer: "再次覆盖旧错误：2021年2月入职", correctsEvidenceId: firstId,
+  }), (error: unknown) => error instanceof ConversationalRectificationError
+    && error.code === "action_conflict");
+  assert.equal(value.cases.get(startActionId)?.row.turnVersion, versionBeforeConflict);
+  assert.equal(value.cases.get(startActionId)?.row.eventEvidence.length, evidenceBeforeConflict);
+});
+
 test("generic date uncertainty does not suppress clear historical evidence", async () => {
   const value = harness();
   await start(value, null);
diff --git a/frontend/tests/conversational-rectification-store.test.ts b/frontend/tests/conversational-rectification-store.test.ts
index 971edc34..f4e36526 100644
--- a/frontend/tests/conversational-rectification-store.test.ts
+++ b/frontend/tests/conversational-rectification-store.test.ts
@@ -288,7 +288,8 @@ test("save, pause, abandon, confirm, and import carry owner/version/action guard
     eventSummary: "换工作",
     dateValue: "2019-07",
     datePrecision: "month" as const,
-    extractionStatus: "clear" as const,
+    extractionStatus: "corrected" as const,
+    correctsEvidenceIds: ["00000000-0000-4000-8000-000000000109"],
   }];
 
   await store.saveTurn({
@@ -731,12 +732,33 @@ test("life-event evidence rejects unknown, blank, and non-boolean durable values
     scoreable: true,
   };
   assert.equal(lifeEventEvidenceSchema.safeParse(evidence).success, true);
+  assert.equal(lifeEventEvidenceSchema.safeParse({
+    ...evidence,
+    extractionStatus: "corrected",
+    correctsEvidenceIds: ["00000000-0000-4000-8000-000000000109"],
+  }).success, true);
+  assert.equal(lifeEventEvidenceSchema.safeParse({
+    ...evidence,
+    extractionStatus: "needs_clarification",
+    scoreable: false,
+    correctsEvidenceIds: ["00000000-0000-4000-8000-000000000109"],
+  }).success, true);
   for (const invalid of [
     { ...evidence, extra: "discard me" },
     { ...evidence, eventSummary: " \t " },
     { ...evidence, dateValue: " \t " },
     { ...evidence, scoreable: null },
     { ...evidence, scoreable: "true" },
+    { ...evidence, extractionStatus: "corrected", correctsEvidenceIds: [] },
+    { ...evidence, correctsEvidenceIds: ["not-a-uuid"] },
+    {
+      ...evidence,
+      extractionStatus: "corrected",
+      correctsEvidenceIds: [
+        "00000000-0000-4000-8000-000000000109",
+        "00000000-0000-4000-8000-000000000110",
+      ],
+    },
   ]) {
     assert.equal(lifeEventEvidenceSchema.safeParse(invalid).success, false);
   }
diff --git a/tests/test_conversational_rectification_contract.py b/tests/test_conversational_rectification_contract.py
index 5bd370f6..fee510fc 100644
--- a/tests/test_conversational_rectification_contract.py
+++ b/tests/test_conversational_rectification_contract.py
@@ -489,6 +489,30 @@ def test_life_event_evidence_is_strictly_validated_before_insert() -> None:
     assert "nullif(item ->> 'datevalue', '')" not in save
 
 
+def test_evidence_correction_lineage_is_bounded_projected_and_validated_atomically() -> None:
+    schema = _normalized(SCHEMA)
+    transitions = _normalized(TRANSITIONS)
+    validator = _function(schema, "conversational_rectification_valid_life_event_evidence")
+    load = _function(transitions, "load_conversational_rectification_case")
+    limits = _function(transitions, "conversational_rectification_case_fits_load_limits")
+    save = _function(transitions, "save_conversational_rectification_turn")
+
+    assert "corrects_evidence_ids uuid[] not null default '{}'::uuid[]" in schema
+    assert "'correctsevidenceids'" in validator
+    assert "jsonb_array_length(p_value -> 'correctsevidenceids') > 1" in validator
+    assert "p_value ->> 'extractionstatus' = 'corrected'" in validator
+    assert "p_value ->> 'extractionstatus' = 'clear'" in validator
+    assert "'correctsevidenceids', pg_catalog.to_jsonb(evidence.corrects_evidence_ids)" in load
+    assert "'correctsevidenceids', pg_catalog.to_jsonb(evidence.corrects_evidence_ids)" in limits
+    assert "corrects_evidence_ids" in save
+    assert "item -> 'correctsevidenceids'" in save
+    assert "pg_catalog.unnest( evidence.corrects_evidence_ids )" in save
+    assert save.index("pg_catalog.unnest( evidence.corrects_evidence_ids )") < save.index(
+        "insert into public.birth_time_rectification_event_evidence"
+    )
+    assert "where evidence.case_id = p_case_id" in save
+
+
 def test_case_mutations_enforce_cumulative_load_limits_under_the_case_lock() -> None:
     transitions = _normalized(TRANSITIONS)
     helper = _function(transitions, "conversational_rectification_case_fits_load_limits")
diff --git a/tests/test_conversational_rectification_postgres_runtime.py b/tests/test_conversational_rectification_postgres_runtime.py
index d3fc23f7..f71d03aa 100644
--- a/tests/test_conversational_rectification_postgres_runtime.py
+++ b/tests/test_conversational_rectification_postgres_runtime.py
@@ -1018,6 +1018,144 @@ def test_save_rejects_invalid_evidence_without_discarding_or_coercing_fields(
     ) == "false"
 
 
+def test_correction_lineage_is_append_only_owner_scoped_and_rejects_retired_targets(
+    pg14_database: PgDatabase,
+) -> None:
+    user_id = "00000000-0000-4000-8000-000000001201"
+    case_id = "00000000-0000-4000-8000-000000001202"
+    first_action = "00000000-0000-4000-8000-000000001203"
+    second_action = "00000000-0000-4000-8000-000000001204"
+    third_action = "00000000-0000-4000-8000-000000001205"
+    rejected_action = "00000000-0000-4000-8000-000000001206"
+    missing_id = "00000000-0000-4000-8000-000000001207"
+    first_id = "00000000-0000-4000-8000-000000001211"
+    second_id = "00000000-0000-4000-8000-000000001212"
+    third_id = "00000000-0000-4000-8000-000000001213"
+    _create_user(pg14_database, user_id)
+    _reserve(pg14_database, user_id, case_id)
+    _create_case(pg14_database, user_id, case_id, _valid_declared_birth_input())
+    _complete(pg14_database, user_id, case_id)
+
+    first = {
+        "id": first_id,
+        "rawText": "2019 年 7 月开始第一份工作",
+        "domain": "career",
+        "eventSummary": "开始第一份工作",
+        "dateValue": "2019-07",
+        "datePrecision": "month",
+        "extractionStatus": "clear",
+        "scoreable": True,
+    }
+    pg14_database.sql(_save_statement(
+        user_id,
+        case_id,
+        0,
+        first_action,
+        [first],
+        turn={
+            **_valid_turn(case_id),
+            "turnVersion": 1,
+            "evidenceRecap": [{
+                "id": first_id,
+                "summary": "开始第一份工作",
+                "dateLabel": "2019-07",
+                "isCorrection": False,
+            }],
+        },
+    ))
+    loaded_first = json.loads(pg14_database.sql(
+        f"select public.load_conversational_rectification_case('{user_id}'::uuid, '{case_id}'::uuid)::text"
+    ))
+    assert loaded_first["event_evidence"][0]["correctsEvidenceIds"] == []
+
+    second = {
+        "id": second_id,
+        "rawText": "更正：2020 年 11 月离职",
+        "domain": "career",
+        "eventSummary": "离职",
+        "dateValue": "2020-11",
+        "datePrecision": "month",
+        "extractionStatus": "corrected",
+        "scoreable": True,
+        "correctsEvidenceIds": [first_id],
+    }
+    pg14_database.sql(_save_statement(
+        user_id,
+        case_id,
+        1,
+        second_action,
+        [second],
+        turn={
+            **_valid_turn(case_id),
+            "turnVersion": 2,
+            "evidenceRecap": [{
+                "id": second_id,
+                "summary": "离职",
+                "dateLabel": "2020-11",
+                "isCorrection": True,
+            }],
+        },
+    ))
+
+    for invalid_target in (missing_id, first_id):
+        rejected = {
+            **second,
+            "id": third_id,
+            "correctsEvidenceIds": [invalid_target],
+        }
+        assert pg14_database.rejects(_save_statement(
+            user_id,
+            case_id,
+            2,
+            rejected_action,
+            [rejected],
+        ))
+    assert pg14_database.sql(
+        f"select turn_version::text from public.birth_time_rectification_cases where id = '{case_id}'::uuid"
+    ) == "2"
+    assert pg14_database.sql(
+        f"select count(*)::text from public.birth_time_rectification_event_evidence where case_id = '{case_id}'::uuid"
+    ) == "2"
+
+    third = {
+        **second,
+        "id": third_id,
+        "rawText": "更正：2021 年 2 月入职",
+        "eventSummary": "入职",
+        "dateValue": "2021-02",
+        "correctsEvidenceIds": [second_id],
+    }
+    pg14_database.sql(_save_statement(
+        user_id,
+        case_id,
+        2,
+        third_action,
+        [third],
+        turn={
+            **_valid_turn(case_id),
+            "turnVersion": 3,
+            "evidenceRecap": [{
+                "id": third_id,
+                "summary": "入职",
+                "dateLabel": "2021-02",
+                "isCorrection": True,
+            }],
+        },
+    ))
+    loaded = json.loads(pg14_database.sql(
+        f"select public.load_conversational_rectification_case('{user_id}'::uuid, '{case_id}'::uuid)::text"
+    ))
+    assert len(loaded["event_evidence"]) == 3
+    assert loaded["event_evidence"][1]["correctsEvidenceIds"] == [first_id]
+    assert loaded["event_evidence"][2]["correctsEvidenceIds"] == [second_id]
+    assert loaded["latest_turn"]["evidenceRecap"] == [{
+        "id": third_id,
+        "summary": "入职",
+        "dateLabel": "2021-02",
+        "isCorrection": True,
+    }]
+
+
 def test_save_rejects_cumulative_evidence_count_before_inserting(
     pg14_database: PgDatabase,
 ) -> None: