diff --git a/crates/utopia-extract/src/lib.rs b/crates/utopia-extract/src/lib.rs
index 428bb07c7..ddcdbff52 100644
--- a/crates/utopia-extract/src/lib.rs
+++ b/crates/utopia-extract/src/lib.rs
@@ -771,6 +771,13 @@ Names:\n\
   \"DeepMind\", \"OpenAI Ireland Ltd\" is not \"OpenAI\", \"Microsoft AI\" is not \
   \"Microsoft\"), a project, programme, team, app or component. Never merge a version into \
   its family or a part into its whole.\n\
+- A document, agreement or filing is cited in many ways and stays one thing through its \
+  amendments: \"the Lease\", \"Lease Agreement\", \"Lease Agreement dated May 16, 2016\" and \
+  \"Lease Agreement dated May 16, 2016, as amended\" are one agreement when their parties and \
+  subject do not contradict each other. When it was signed and that it was amended describe \
+  the agreement; they do not make a second one. Each amendment is a document of its own, and \
+  an agreement for other premises, another phase or other parties (\"Phase 2 Lease\") is a \
+  different agreement.\n\
 - A phrase that merely contains a name is not that name: \"Sam Altman's efforts\", \
   \"psychological abuse from Sam Altman\", \"share sale led by Thrive Capital\", \"leaked \
   letter from the National Data Guardian\", \"ChatGPT played a role in the campaign\", \"a \
diff --git a/crates/utopia-server/src/governance.rs b/crates/utopia-server/src/governance.rs
index 3e4253822..367250fca 100644
--- a/crates/utopia-server/src/governance.rs
+++ b/crates/utopia-server/src/governance.rs
@@ -370,7 +370,10 @@ fn pair_of(item: &ReviewItem, p: &Precedents) -> utopia_extract::AdjudicationPai
 }
 
 /// 第一层没定（没判决、置信度不到线），或者同名、大类不冲突、模型却说不同——那是它
-/// 最爱错的一种：都让第二层带着全部事实与原文再看一遍。硬规则拦下的不进：再看也不会改规则
+/// 最爱错的一种：都让第二层带着全部事实与原文再看一遍。名字的形状让人起疑的合并（版本
+/// 尾巴、含着名字的一长串）也再看：名字只能让人起疑，定不了是不是一个东西——「Lease
+/// Agreement dated May 16, 2016, as amended」与「Lease Agreement」形状上是一句话，事实上
+/// 是同一份租约。类型冲突这条硬规则拦下的不进：再看也不会改规则
 fn wants_second_look(item: &ReviewItem, p: &Precedents, look: &Look) -> bool {
     let types_conflict = gov::types_conflict(
         item.left.type_label.as_deref(),
@@ -381,12 +384,38 @@ fn wants_second_look(item: &ReviewItem, p: &Precedents, look: &Look) -> bool {
         && !types_conflict
         && look.same == Some(false)
         && look.calls == 0;
+    let doubted_merge =
+        name_doubts(shape) && !types_conflict && look.same == Some(true) && look.calls == 0;
     ((gov::gate(look.same, look.conf, types_conflict, shape, p) == Gate::Propose
         && look.uncertain())
-        || doubted_split)
+        || doubted_split
+        || doubted_merge)
         && p.reverts.is_empty()
 }
 
+/// 名字形状让人起疑的合并：版本尾巴，或含着另一个名字的一长串
+fn name_doubts(shape: gov::NameShape) -> bool {
+    matches!(shape, gov::NameShape::Version | gov::NameShape::Phrase)
+}
+
+/// 闸门看的名字形状。第二层带着两边的事实与原文看过、仍说是同一个的，形状的疑点已经由
+/// 证据答过了，不再按形状拦——拦的只剩没看过证据的第一层
+fn shape_for_gate(item: &ReviewItem, look: &Look) -> gov::NameShape {
+    settled_shape(
+        gov::name_shape(&item.left.name, &item.right.name),
+        look.calls > 0,
+    )
+}
+
+/// 形状的疑点在第二层读过证据之后就答完了
+fn settled_shape(shape: gov::NameShape, evidence_read: bool) -> gov::NameShape {
+    if evidence_read && name_doubts(shape) {
+        gov::NameShape::Unrelated
+    } else {
+        shape
+    }
+}
+
 /// 裁决器的入口（0028）：治理关着，攒批判不定的对也带工具再看一遍——同一个循环、
 /// 同一份预算。每次调用一个 run_id：一次裁决任务就是一次 run
 pub(crate) async fn look_again(
@@ -446,7 +475,7 @@ async fn apply(ctx: &Ctx<'_>, item: &ReviewItem, p: &Precedents, look: Look) ->
         item.left.type_label.as_deref(),
         item.right.type_label.as_deref(),
     );
-    let shape = gov::name_shape(&item.left.name, &item.right.name);
+    let shape = shape_for_gate(item, &look);
 
     let action = look.action();
     let precedents = gov::precedents_json(p);
@@ -908,3 +937,28 @@ pub async fn after_human_decision(
         _ => {}
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn a_doubtful_shape_stops_counting_once_the_evidence_was_read() {
+        let lease = gov::name_shape(
+            "Lease Agreement dated May 16, 2016, as amended",
+            "Lease Agreement",
+        );
+        assert!(name_doubts(lease), "{lease:?}");
+        assert_eq!(settled_shape(lease, false), lease);
+        assert_eq!(settled_shape(lease, true), gov::NameShape::Unrelated);
+
+        let version = gov::name_shape("Claude Mythos 5", "Claude Mythos");
+        assert!(name_doubts(version), "{version:?}");
+        assert_eq!(settled_shape(version, false), version);
+
+        // 形状本来就不起疑的，读没读过证据都照旧
+        let same = gov::name_shape("OpenAI", "OpenAI");
+        assert!(!name_doubts(same));
+        assert_eq!(settled_shape(same, true), same);
+    }
+}
diff --git a/docs/decisions/0025-governance-reads-the-ledger-before-it-decides.md b/docs/decisions/0025-governance-reads-the-ledger-before-it-decides.md
index 529696226..556f37ba3 100644
--- a/docs/decisions/0025-governance-reads-the-ledger-before-it-decides.md
+++ b/docs/decisions/0025-governance-reads-the-ledger-before-it-decides.md
@@ -1,6 +1,6 @@
 # 0025 · Governance reads the ledger before it decides
 
-- **Status**: decision 3 revised 2026-09-07 (three clusters a round, #458) · decision 10 added 2026-09-07 (identity rules, name shapes, type families, the labeled set and `govern.mjs`) · decision 4 revised 2026-09-06 (the agent's own confidence decides, history moves the bar or blocks) · cut 1 implemented (#434) · migration 0035 adds `knowledge_bases.governance` and `agent_decisions`; `governance` in the store holds the precedent families, the first-in-first-out queue with its clusters, the gate and the table; the `govern` job in the server reads the switch, calls the model with precedents and applies or proposes; `?queue=agent`, `ReviewCounts.agent` and `POST /kbs/{id}/review/agent/{decision_id}` on the API · cut 3 (UI, #437) implemented: the switch in base settings, the Agent queue with its rows and answers, the proposal chip on a duplicate card whose Merge / Keep answers the proposal, the Agent section on the Overview · cut 2 implemented: migration 0036 adds `question`, `trace` and `calls` to `agent_decisions`; `governor` in the extract crate holds the tools, the opening and the step reader; `investigate` in the server job runs the loop for pairs the batch could not settle (decision 8) · cut 4 implemented: migration 0037 adds `knowledge_bases.governance_since`; `fuse` in the server job turns the switch off after two reverts of the agent's merges since it was turned on, within seven days, raises `governance.tripped` and writes the ledger (decision 9) · [0026](0026-a-decision-records-why.md) gives a decision its stated reason, hands the batch adjudicator the same precedents and quotes the reason into them · [0028](0028-the-adjudicator-looks-before-it-asks.md) opens the loop of cut 2 to the adjudicator with governance off and gives it a `consequences` tool
+- **Status**: decision 10 revised 2026-09-14 (a shape that raises doubt goes to the second look, and what that look finds is applied, #700) · decision 3 revised 2026-09-07 (three clusters a round, #458) · decision 10 added 2026-09-07 (identity rules, name shapes, type families, the labeled set and `govern.mjs`) · decision 4 revised 2026-09-06 (the agent's own confidence decides, history moves the bar or blocks) · cut 1 implemented (#434) · migration 0035 adds `knowledge_bases.governance` and `agent_decisions`; `governance` in the store holds the precedent families, the first-in-first-out queue with its clusters, the gate and the table; the `govern` job in the server reads the switch, calls the model with precedents and applies or proposes; `?queue=agent`, `ReviewCounts.agent` and `POST /kbs/{id}/review/agent/{decision_id}` on the API · cut 3 (UI, #437) implemented: the switch in base settings, the Agent queue with its rows and answers, the proposal chip on a duplicate card whose Merge / Keep answers the proposal, the Agent section on the Overview · cut 2 implemented: migration 0036 adds `question`, `trace` and `calls` to `agent_decisions`; `governor` in the extract crate holds the tools, the opening and the step reader; `investigate` in the server job runs the loop for pairs the batch could not settle (decision 8) · cut 4 implemented: migration 0037 adds `knowledge_bases.governance_since`; `fuse` in the server job turns the switch off after two reverts of the agent's merges since it was turned on, within seven days, raises `governance.tripped` and writes the ledger (decision 9) · [0026](0026-a-decision-records-why.md) gives a decision its stated reason, hands the batch adjudicator the same precedents and quotes the reason into them · [0028](0028-the-adjudicator-looks-before-it-asks.md) opens the loop of cut 2 to the adjudicator with governance off and gives it a `consequences` tool
 - **Written**: 2026-09-06 (conventions in the [README](README.md))
 - **Related**: [0016](0016-close-the-open-seams-before-cutting-new-ones.md) C2 gave a base its first automation switch, `auto_type_resolution`, and this record copies its shape; #428 asked for bulk and automatic handling of same-name pairs and got the batch path (#429, #430) this builds on; [0020](0020-an-auditor-reads-it-without-us.md) made the ledger complete enough to be read back; [0015](0015-recording-a-sentence-is-not-asserting-a-fact.md) keeps a person's own sentences out of any machine's reach.
 
@@ -75,7 +75,7 @@ Both revert paths count: the Agent queue's Revert and the Merges page's Revert o
 
 The first big-file run (decision 4's revision note) showed two things at once: a fresh base needs the agent to decide on its own, and what the model decides on its own is wrong in patterned ways. So on 2026-09-07 every pair the resolver raised on one governed import of eight interlocking articles (615 pairs, 411 distinct name pairs) was labeled by hand from each side's facts and source passages: `same`, `different`, or `unknown` where no reader could tell either (a bare "judge"). The set lives in `scripts/bench/truth/ai-timeline.duplicates.json`, keyed by the two names, and `scripts/bench/govern.mjs` resets a base's agent decisions without touching the ledger, runs the governor, and scores: decided on its own, agreed with the labels, wrong merges, wrong keeps, left for people.
 
-The patterns became rules, written once (`IDENTITY_RULES`) and read by both prompts: a version, edition, tier, division, subsidiary or legal entity added at the end of a name is a different thing; a phrase that contains a name is not the name; a list is not its members; a common noun is not a proper name; a qualifier dropped from the front, a corporate suffix, a surname for a full name, a parenthetical acronym are the same thing; a different type label, a missing fact, a different partner or role is not a contradiction. Two of the rules are also mechanical, so the model cannot talk its way past them: `name_shape` classifies a pair as identical, abbreviation, suffix, version, extension, phrase or unrelated, and a `same` verdict on a version or a phrase is never applied on its own; `type_family` folds the extractor's labels into person, organization, place and event, and only a difference across families counts as a conflict. Identical names whose labels fall in one family are shown to the model with one label, since the label difference was the reason it gave most often for a wrong split; and a `different` verdict on identical, compatible names goes to the tool loop for a second look before it is applied.
+The patterns became rules, written once (`IDENTITY_RULES`) and read by both prompts: a version, edition, tier, division, subsidiary or legal entity added at the end of a name is a different thing; a phrase that contains a name is not the name; a list is not its members; a common noun is not a proper name; a qualifier dropped from the front, a corporate suffix, a surname for a full name, a parenthetical acronym are the same thing; a different type label, a missing fact, a different partner or role is not a contradiction. Two of the rules are also mechanical, so the model cannot talk its way past them: `name_shape` classifies a pair as identical, abbreviation, suffix, version, extension, phrase or unrelated, and a `same` verdict on a version or a phrase is never applied on the first look alone; `type_family` folds the extractor's labels into person, organization, place and event, and only a difference across families counts as a conflict. Identical names whose labels fall in one family are shown to the model with one label, since the label difference was the reason it gave most often for a wrong split; and a `different` verdict on identical, compatible names goes to the tool loop for a second look before it is applied. Revised 2026-09-14 (#700): a `same` verdict on a version or a phrase goes to that second look too, and once the loop has read both sides' facts and text and still says `same`, the gate reads the pair without its shape. The shape raises a doubt, and only the evidence can settle it: "Lease Agreement dated May 16, 2016, as amended" and "Lease Agreement" are a phrase by shape and one lease by their facts. The identity rules gain the matching sentence: a document, agreement or filing stays one thing through its amendments, each amendment is its own document, and an agreement for other premises, another phase or other parties is a different one. On the labeled set, both runs reset and re-decided on the same extraction: agreement 96.3% → 98.1%, wrong merges 2 → 1, wrong keeps 4 → 2, pairs left for people 30 → 34.
 
 The same 589 pairs, reset and re-run after each change:
 
diff --git a/scripts/bench/govern.mjs b/scripts/bench/govern.mjs
index c8ffdefce..8ab48ff5b 100644
--- a/scripts/bench/govern.mjs
+++ b/scripts/bench/govern.mjs
@@ -84,7 +84,10 @@ async function fresh() {
     await api("POST", `/api/v1/kbs/${kb}/ingest`, body);
   }
   await until(async () => {
-    const done = num(`SELECT count(*) FROM documents WHERE kb_id='${kb}' AND graph_status IN ('done','failed')`);
+    // 失败了但抽取任务还排着的（本体索引没嵌完、延后重试）不算完
+    const done = num(`SELECT count(*) FROM documents d WHERE d.kb_id='${kb}' AND (d.graph_status = 'done'
+      OR (d.graph_status = 'failed' AND NOT EXISTS (SELECT 1 FROM jobs j WHERE j.kind = 'extract_document'
+          AND j.status IN ('queued','running') AND j.payload->>'document_id' = d.id::text)))`);
     if (done >= docs.length) return true;
     const chunks = num(`SELECT count(*) FROM chunks WHERE kb_id='${kb}' AND extracted_at IS NOT NULL`);
     log(`  抽取 ${chunks} 块 / ${done} 篇完成`);