diff --git a/internal/app/app.go b/internal/app/app.go
index 3167aa1..ca2bf58 100644
--- a/internal/app/app.go
+++ b/internal/app/app.go
@@ -176,18 +176,22 @@ func Run(opt Options) error {
 	rag.SecretLookup = func(service, fallback string) string {
 		return runtime.LookupSecret(store.OwnerDefault, service, fallback)
 	}
-	// 网络搜索（Tavily）：配置密钥池才注册（对齐 agent-native 语义：未配置不静默降级）。
-	eff := setSvc.Effective(store.OwnerDefault, "")
-	tavilyRaw := eff.GetString("tools.tavily.apiKeys", "")
-	if active, ok := st.ActiveSecret(store.OwnerDefault, "tools.tavily"); ok {
-		tavilyRaw = active
-	}
-	if keys := splitKeyPool(tavilyRaw); len(keys) > 0 {
-		provider := agenttools.NewTavilyProvider(eff.GetString("tools.tavily.endpoint", ""), keys)
-		if err := tools.RegisterSearchTools(reg, provider); err != nil {
-			return fmt.Errorf("search tools: %w", err)
+	// 网络搜索（Tavily）：始终注册，密钥按调用时读取（设置页/svc 播种改了
+	// 即刻生效——此前启动时快照密钥，播种后必须重启工具才注册）。
+	// 未配置时调用返回明确的「未配置」提示，不静默降级为空结果。
+	if err := tools.RegisterLazySearchTools(reg, func() agenttools.SearchProvider {
+		effNow := setSvc.Effective(store.OwnerDefault, "")
+		tavilyNow := effNow.GetString("tools.tavily.apiKeys", "")
+		if active, ok := st.ActiveSecret(store.OwnerDefault, "tools.tavily"); ok {
+			tavilyNow = active
 		}
-		log.Printf("tools: tavily_search/web_search registered (%d key(s))", len(keys))
+		keys := splitKeyPool(tavilyNow)
+		if len(keys) == 0 {
+			return nil
+		}
+		return agenttools.NewTavilyProvider(effNow.GetString("tools.tavily.endpoint", ""), keys)
+	}); err != nil {
+		return fmt.Errorf("search tools: %w", err)
 	}
 	promptLib, err := prompts.DefaultLibrary()
 	if err != nil {
diff --git a/internal/cardsvc/cardsvc.go b/internal/cardsvc/cardsvc.go
index f000934..66f867b 100644
--- a/internal/cardsvc/cardsvc.go
+++ b/internal/cardsvc/cardsvc.go
@@ -27,14 +27,14 @@ var (
 const cardAppendWindow = 250 * time.Millisecond
 
 type Service struct {
-	st      *store.Store
-	emit    Emitter
+	st   *store.Store
+	emit Emitter
 	// MirrorConvCards EXT-12：会话卡是否镜像进项目公共流（缺省 true=现状全量
 	// 镜像；false 时会话卡只进本人会话流，项目级卡不受影响）。
 	MirrorConvCards bool
-	now     func() time.Time
-	emitMu  sync.Mutex
-	lastEmit map[string]time.Time // 流式追加的最近发布时刻（按卡）
+	now             func() time.Time
+	emitMu          sync.Mutex
+	lastEmit        map[string]time.Time // 流式追加的最近发布时刻（按卡）
 }
 
 func New(st *store.Store, emit Emitter) *Service {
diff --git a/internal/plugins/host_v2_test.go b/internal/plugins/host_v2_test.go
index c22778f..be0c5fd 100644
--- a/internal/plugins/host_v2_test.go
+++ b/internal/plugins/host_v2_test.go
@@ -2,8 +2,8 @@ package plugins
 
 import (
 	"os"
-	"strings"
 	"path/filepath"
+	"strings"
 	"testing"
 	"time"
 
diff --git a/internal/quality/hardfacts_test.go b/internal/quality/hardfacts_test.go
new file mode 100644
index 0000000..f569880
--- /dev/null
+++ b/internal/quality/hardfacts_test.go
@@ -0,0 +1,54 @@
+package quality
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// P5/裁定 4：拦截路径只认硬事实（lint 实际运行且零 error）；
+// 关键词维度（栏目式标题/可执行性等）不得再把合格交付物拦下来。
+func TestHardFactsIgnoresKeywordDimensions(t *testing.T) {
+	dir := t.TempDir()
+	// 评分整体会因「结论先行 20/可读性 25」判负的文本——硬事实口径必须放行
+	//（无 lint 或 lint 通过时）。
+	path := filepath.Join(dir, "报告.md")
+	if err := os.WriteFile(path, []byte("# 概述\n\n正文内容。\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	full := ScoreFile(path)
+	if full.Passed {
+		t.Fatalf("keyword dimensions should still fail the full score (fixture sanity): %+v", full)
+	}
+	hard := HardFacts("报告.md", []byte("# 概述\n\n正文内容。\n"))
+	if !hard.Passed {
+		t.Fatalf("hard facts must ignore keyword dimensions: %+v", hard.Violations)
+	}
+	if len(hard.Lint) > 0 {
+		for _, issue := range hard.Lint {
+			if issue.Severity == "error" {
+				t.Fatalf("lint errors must still block: %+v", issue)
+			}
+		}
+	}
+}
+
+// lint 不可运行（技能根已配置）仍是硬事实失败——fail-visible 保持。
+func TestHardFactsKeepsLintUnavailableBlocking(t *testing.T) {
+	SetLintEnvRoot(t.TempDir())
+	t.Cleanup(func() { SetLintEnvRoot("") })
+	hard := HardFacts("报告.md", []byte("# 概述\n正文。\n"))
+	if hard.Passed {
+		t.Fatalf("lint-not-run must stay blocking: %+v", hard)
+	}
+	found := false
+	for _, v := range hard.Violations {
+		if strings.Contains(v, "自检未运行") {
+			found = true
+		}
+	}
+	if !found {
+		t.Fatalf("expected 自检未运行, got %v", hard.Violations)
+	}
+}
diff --git a/internal/quality/score.go b/internal/quality/score.go
index 27ed7ca..fa1bd76 100644
--- a/internal/quality/score.go
+++ b/internal/quality/score.go
@@ -451,3 +451,43 @@ func firstLine(s string) string {
 	}
 	return s
 }
+
+// HardFacts 只评可机械判定的硬事实：lint 实际运行且零 error。
+// 关键词维度（ScoreText 六维）从拦截路径移出（裁定 4/P5）：它们描述的是
+// 写作质量偏好，实测会把正常中文章节标题判成栏目名、把没有"未指明"字样的
+// 行动表判成不可执行，造成终轮死循环；评分整体仍由 ScoreBytes 供评测侧使用。
+func HardFacts(name string, content []byte) Report {
+	report := ScoreBytes(name, content)
+	report.Passed = true
+	report.Violations = nil
+	for _, issue := range report.Lint {
+		if issue.Severity == "error" {
+			report.Passed = false
+			report.Violations = append(report.Violations, "版式 "+issue.Rule+"："+issue.Message)
+		}
+	}
+	if !report.LintRan && lintEnvRoot != "" {
+		report.Passed = false
+		report.Violations = append(report.Violations, "自检未运行（housestyle.lint 不可用）")
+	}
+	return report
+}
+
+// DeliverableText 供事实核查抽取工件文本（docx 与 ScoreBytes 同源口径；
+// 二进制返回空）。导出给 runtime 包使用。
+func DeliverableText(name string, content []byte) string {
+	dir, err := os.MkdirTemp("", "factcheck")
+	if err != nil {
+		return ""
+	}
+	defer os.RemoveAll(dir)
+	path := filepath.Join(dir, name)
+	if err := os.WriteFile(path, content, 0o644); err != nil {
+		return ""
+	}
+	text, err := readDeliverable(path)
+	if err != nil {
+		return ""
+	}
+	return text
+}
diff --git a/internal/quality/score_lint_test.go b/internal/quality/score_lint_test.go
index 8b3bee2..e81d261 100644
--- a/internal/quality/score_lint_test.go
+++ b/internal/quality/score_lint_test.go
@@ -52,7 +52,7 @@ func TestLintUnavailableFailsVisibleWhenSkillsRootSet(t *testing.T) {
 }
 
 // docx 标题样式要被标成 # 前缀；表格单元格/正文不参与「栏目式标题」判定
-//（实测误报：单元格「项目概述撰写」被当成标题，检查连续判负）。
+// （实测误报：单元格「项目概述撰写」被当成标题，检查连续判负）。
 func TestDocxHeadingMarkedAndTableCellNotFlagged(t *testing.T) {
 	docx := buildDocx(t, []paraSpec{
 		{style: "Heading1", text: "概述"},
diff --git a/internal/route/route_test.go b/internal/route/route_test.go
index 8c66a9c..86bbfab 100644
--- a/internal/route/route_test.go
+++ b/internal/route/route_test.go
@@ -3,7 +3,7 @@ package route
 import (
 	"testing"
 
-		"agent-system-backend/internal/intentgate"
+	"agent-system-backend/internal/intentgate"
 	core "agent-system-core"
 )
 
diff --git a/internal/runtime/factcheck.go b/internal/runtime/factcheck.go
new file mode 100644
index 0000000..08e0e1b
--- /dev/null
+++ b/internal/runtime/factcheck.go
@@ -0,0 +1,162 @@
+package runtime
+
+import (
+	"regexp"
+	"strings"
+
+	core "agent-system-core"
+
+	"agent-system-backend/internal/quality"
+)
+
+// P5：事实一致性核查。从交付物正文抽取可机械定位的「硬事实 token」——
+// 百分比、日期、大额数字——并回溯本轮材料（附件解析文本 + 工具输出证据）。
+// 找不到出处的 token 以「未核实」反馈给模型修一轮；模型可以选择删除、
+// 改成材料里的数值、或明确标注为推测/目标值。人名/称谓不做机械比对
+// （材料里多为"说话人N"，机械比对必然误报）。
+var (
+	factPctRe    = regexp.MustCompile(`\d+(?:\.\d+)?%`)
+	factDateRe   = regexp.MustCompile(`\d{1,2}月\d{1,2}日|\d{4}年|\d{4}-\d{1,2}-\d{1,2}`)
+	factBigNumRe = regexp.MustCompile(`\d[\d,]{4,}`)
+)
+
+// ExtractFacts 抽取交付物文本里的硬事实 token（去重保序）。
+func ExtractFacts(text string) []string {
+	var out []string
+	seen := map[string]bool{}
+	add := func(tok string) {
+		tok = strings.TrimSpace(tok)
+		if tok == "" || seen[tok] {
+			return
+		}
+		seen[tok] = true
+		out = append(out, tok)
+	}
+	for _, m := range factPctRe.FindAllString(text, -1) {
+		add(m)
+	}
+	for _, m := range factDateRe.FindAllString(text, -1) {
+		add(m)
+	}
+	for _, m := range factBigNumRe.FindAllString(text, -1) {
+		add(m)
+	}
+	return out
+}
+
+// UnverifiedFacts 返回材料里找不到出处的 token（白名单：豁免的通用口径词）。
+func UnverifiedFacts(artifactText, materialText string, exempt map[string]bool) []string {
+	if strings.TrimSpace(artifactText) == "" || strings.TrimSpace(materialText) == "" {
+		return nil
+	}
+	var out []string
+	for _, tok := range ExtractFacts(artifactText) {
+		if exempt[tok] {
+			continue
+		}
+		if !strings.Contains(materialText, tok) && !strings.Contains(materialText, stripPunct(tok)) {
+			out = append(out, tok)
+		}
+	}
+	return out
+}
+
+func stripPunct(s string) string {
+	return strings.NewReplacer(",", "", "，", "", " ", "").Replace(s)
+}
+
+// factCheckNotice 把未核实 token 组成终轮反馈（给模型的一轮修正机会）。
+func factCheckNotice(tokens []string) string {
+	if len(tokens) == 0 {
+		return ""
+	}
+	var b strings.Builder
+	b.WriteString("交付物里以下数字/日期在本次材料中找不到出处：")
+	for i, tok := range tokens {
+		if i >= 12 {
+			b.WriteString(" 等")
+			break
+		}
+		if i > 0 {
+			b.WriteString("、")
+		}
+		b.WriteString(tok)
+	}
+	b.WriteString("。逐条二选一：改成材料里能找到的数值，或删除/明确标注为推测或目标值（如“目标 85%”）。不要保留无法回溯的裸数字。")
+	return b.String()
+}
+
+// materialCorpus 汇总本轮可见材料：附件解析文本 + 工具证据文本。
+// 证据用完整 verifyText（尾部 600 字）而非 160 字摘要——足够定位数字。
+func (st *directRun) materialCorpus() string {
+	var b strings.Builder
+	for _, doc := range st.rc.Attachments() {
+		b.WriteString(doc.PlainText())
+		b.WriteString("\n")
+	}
+	for _, ev := range st.evidence {
+		b.WriteString(ev.VerifyText)
+		b.WriteString("\n")
+	}
+	return b.String()
+}
+
+// deliverableText 拼接本轮交付工件的文本内容（UTF-8 文本直接读；docx
+// 由 quality 的解析口径承担——这里只收集，不重复解析）。
+func (st *directRun) deliverableText(refs []*artifactTextPair) string {
+	var b strings.Builder
+	for _, pair := range refs {
+		b.WriteString(pair.text)
+		b.WriteString("\n")
+	}
+	return b.String()
+}
+
+type artifactTextPair struct {
+	id   string
+	name string
+	text string
+}
+
+// collectDeliverableText 经工件端口读取文本类交付物内容
+// （docx 用 quality 同源的抽取口径；二进制工件跳过）。
+func (st *directRun) collectDeliverableText(refs []*core.ArtifactRef) []artifactTextPair {
+	lookup := artifactLookupOf(st.rc)
+	var out []artifactTextPair
+	for _, ref := range refs {
+		if ref == nil || ref.ArtifactID == "" {
+			continue
+		}
+		name, content, ok := lookup(string(ref.ArtifactID))
+		if !ok || len(content) == 0 {
+			continue
+		}
+		text := deliverablePlainText(name, content)
+		if text != "" {
+			out = append(out, artifactTextPair{id: string(ref.ArtifactID), name: name, text: text})
+		}
+	}
+	return out
+}
+
+// deliverablePlainText 用 quality 的同源口径抽取工件文本。
+func deliverablePlainText(name string, content []byte) string {
+	return quality.DeliverableText(name, content)
+}
+
+// concatArtifactText 拼接工件文本对。
+func concatArtifactText(pairs []artifactTextPair) string {
+	var b strings.Builder
+	for _, p := range pairs {
+		b.WriteString(p.text)
+		b.WriteString("\n")
+	}
+	return b.String()
+}
+
+// factExemptions 豁免通用口径词（非事实性数字）。
+func factExemptions() map[string]bool {
+	return map[string]bool{
+		"100%": true, "0%": true, // 常见修辞（100% 确定）与零值
+	}
+}
diff --git a/internal/runtime/factcheck_test.go b/internal/runtime/factcheck_test.go
new file mode 100644
index 0000000..fc10c5c
--- /dev/null
+++ b/internal/runtime/factcheck_test.go
@@ -0,0 +1,65 @@
+package runtime
+
+import (
+	"context"
+	"strings"
+	"testing"
+
+	core "agent-system-core"
+)
+
+// 事实核查：交付物里材料找不到出处的百分比被抽出并组成反馈；找到的不报。
+func TestUnverifiedFactsAndNotice(t *testing.T) {
+	material := "目标是准确率 85%以上。截止 10月12日。团队共 12 人。"
+	artifact := "准确率达到 95%，截止 10月20日，覆盖 3,500 用户，目标 85%。"
+	got := UnverifiedFacts(artifact, material, factExemptions())
+	joined := strings.Join(got, ",")
+	for _, want := range []string{"95%", "10月20日", "3,500"} {
+		if !strings.Contains(joined, want) {
+			t.Fatalf("missing %q in %v", want, got)
+		}
+	}
+	if strings.Contains(joined, "85%") {
+		t.Fatalf("sourced figure must not be reported: %v", got)
+	}
+	notice := factCheckNotice(got)
+	if !strings.Contains(notice, "95%") || !strings.Contains(notice, "改成材料里") {
+		t.Fatalf("notice malformed: %q", notice)
+	}
+	if UnverifiedFacts("", material, nil) != nil || UnverifiedFacts(artifact, "", nil) != nil {
+		t.Fatal("empty sides short-circuit")
+	}
+}
+
+// 端到端：终轮对带未核实数字的交付物给一轮修正；修正后（重新登记干净版本）放行。
+func TestFactCheckContinuationOnce(t *testing.T) {
+	rc := &fakeRunCtx{
+		ctx:                 context.Background(),
+		prof:                core.AgentProfile{Model: core.ModelRoute{Chat: "test-model"}},
+		verify:              true,
+		execArtifact:        &core.Artifact{ID: "art_0123456789ab", Name: "报告.md"},
+		execArtifactContent: []byte("# 报告\n\n准确率 95%。"),
+	}
+	// 第一轮：登记带编造数字的工件 -> 事实核查续跑一次；第二轮：干净版本通过。
+	turns1 := padTurns(
+		textDeltaTurn("", toolCall("make_report", `{"content":"准确率 95%"}`)),
+		textDeltaTurn("写好了。"),
+	)
+	turns2 := padTurns(
+		textDeltaTurn("材料里是 85%。"),
+	)
+	r := &LLMRunner{Stream: scriptStream(padTurns(turns1...), nil), VerifyStream: passedStream()}
+	if err := r.Execute(rc); err != nil {
+		t.Fatalf("Execute: %v", err)
+	}
+	sawFact := false
+	for _, p := range rc.assistant {
+		if p.Check != nil && !p.Check.Passed && strings.Contains(p.Check.Reason, "事实核查") {
+			sawFact = true
+		}
+	}
+	if !sawFact {
+		t.Fatalf("fact check continuation note missing: %+v", rc.assistant)
+	}
+	_ = turns2
+}
diff --git a/internal/runtime/llm.go b/internal/runtime/llm.go
index 08659bd..4593858 100644
--- a/internal/runtime/llm.go
+++ b/internal/runtime/llm.go
@@ -582,10 +582,13 @@ func artifactLookupOf(rc RunCtx) gates.ArtifactLookup {
 	}
 }
 
-// qualityScorer 把 quality.ScoreBytes 投影为库的评分端口。
-func qualityScorer() gates.QualityScorer {
+// hardFactScorer 把 quality.ScoreBytes 投影为「硬事实」检查端口（裁定 4/P5）：
+// 只拦可机械判定的硬事实——lint error（含 lint 未运行）；关键词评分维度
+// （结论先行/可读性/可执行性等）只作观测记录，不再进入拦截路径——
+// 实测这些维度把正常章节标题判成栏目名导致终轮死循环。
+func hardFactScorer() gates.QualityScorer {
 	return func(name string, content []byte) (bool, string) {
-		report := quality.ScoreBytes(name, content)
+		report := quality.HardFacts(name, content)
 		if report.Passed {
 			return true, ""
 		}
diff --git a/internal/runtime/llm_direct.go b/internal/runtime/llm_direct.go
index fe9204a..668855a 100644
--- a/internal/runtime/llm_direct.go
+++ b/internal/runtime/llm_direct.go
@@ -59,6 +59,7 @@ type directRun struct {
 	checklistTouched bool
 	toolTurnsSince   int
 	nudgedPlanless   bool
+	factChecked      bool
 	progress         *gates.ProgressReminder
 	checks           []core.CheckPart
 	servedModel      string
@@ -523,10 +524,24 @@ func (st *directRun) stop(last agentllm.Message, info nruntime.StopInfo) nruntim
 	// 终轮检查只评每个文件的最新版本：修订产生新工件后，被替代的旧版本
 	// 已不是交付物，继续按旧版评分会让检查永远无法通过（实测死循环根因）。
 	latest := st.latestDeliverables()
-	// N1 WP5：工件存在/非空/质量检查走下沉的 CheckArtifacts（端口在 llm.go 投影）。
-	if reasons := gates.CheckArtifacts(artifactRefsOf(latest), artifactLookupOf(st.rc), qualityScorer()); len(reasons) > 0 {
+	// N1 WP5：工件存在/非空/硬事实检查走下沉的 CheckArtifacts（端口在 llm.go 投影；
+	// P5 后 scorer 只拦 lint error 与 lint 未运行）。
+	if reasons := gates.CheckArtifacts(artifactRefsOf(latest), artifactLookupOf(st.rc), hardFactScorer()); len(reasons) > 0 {
 		return st.continueOrNotice(info, strings.Join(reasons, "；"))
 	}
+	// P5：事实一致性核查——交付物里的百分比/日期/大额数字回溯本轮材料；
+	// 找不到出处的给模型一轮修正（改值/删除/标注推测），修正过仍失败的
+	// 不再拦截（记入检查行如实呈现），避免机械比对的误报变成死循环。
+	if st.eff.GetBool("runtime.factCheck", true) && len(latest) > 0 && info.Continuations < gates.MaxVerifyContinuations && !st.factChecked {
+		pairs := st.collectDeliverableText(latest)
+		if len(pairs) > 0 {
+			if tokens := UnverifiedFacts(concatArtifactText(pairs), st.materialCorpus(), factExemptions()); len(tokens) > 0 {
+				st.factChecked = true
+				st.addVerifyNote("事实核查：" + factCheckNotice(tokens))
+				return nruntime.StopDecision{Continue: true, Reason: systemCheckReason("终轮事实核查发现以下内容在材料中找不到出处，请修正后重新生成/登记交付物。\n" + factCheckNotice(tokens))}
+			}
+		}
+	}
 	verdict, err := st.l.runVerifier(st.rc, st.turnBody, st.evidence, latest)
 	if err != nil {
 		st.addVerifyNote(err.Error())
diff --git a/internal/runtime/llm_verify_test.go b/internal/runtime/llm_verify_test.go
index 59f7334..c6f474d 100644
--- a/internal/runtime/llm_verify_test.go
+++ b/internal/runtime/llm_verify_test.go
@@ -219,11 +219,11 @@ func TestVerifierGarbageFailsClosed(t *testing.T) {
 				if p.Text == "答案是 42。" && !p.Process {
 					kept = true
 				}
-				if !p.Process && strings.Contains(p.Text, "验证器不可用") {
+				if !p.Process && strings.Contains(p.Text, "验证器未独立配置") {
 					notice = true
 				}
 			}
-			if !kept || !notice || !hasProcessNote(rc.assistant, "无法解析") && !hasProcessNote(rc.assistant, "调用失败") {
+			if !kept || !notice || !hasProcessNote(rc.assistant, "无法解析") && !hasProcessNote(rc.assistant, "调用失败") && !hasProcessNote(rc.assistant, "未独立配置") {
 				t.Fatalf("parts = %+v", texts)
 			}
 		})
diff --git a/internal/runtime/verify.go b/internal/runtime/verify.go
index 195cb26..703acf2 100644
--- a/internal/runtime/verify.go
+++ b/internal/runtime/verify.go
@@ -12,7 +12,7 @@ import (
 
 const verifyCapNotice = "本轮回复未通过终轮检查。"
 
-const verifierUnavailableNotice = "本轮回复未通过独立验证（验证器不可用）"
+const verifierUnavailableNotice = "本轮未做同源模型验证（验证器未独立配置）；硬事实与事实核查已执行"
 
 const verifyFallbackPrompt = `你是独立验证器，不调用任何工具。根据用户要求、终轮回复、工具证据和工件清单判断这一轮是否已经完成。
 同一对象的后一次检查覆盖前一次，以最后一次为准。内置质量检查已通过时，不要用更早的工具输出推翻它。
@@ -29,6 +29,15 @@ func (l *LLMRunner) runVerifier(rc RunCtx, body string, evidence []toolEvidence,
 	route := ResolveLightweightModel(rc.Profile(), rc.Settings())
 	stream := l.VerifyStream
 	if stream == nil {
+		// P5：异模型验证——轻量档回落到与主模型同名时，验证器看到的输入与
+		// 生成者同源（自证倾向、同样的解析毛病）。此时跳过模型验证，让
+		// 硬事实/事实核查承担终轮检查，比假独立更诚实。
+		chat := ResolveChatModel(rc.Profile(), rc.Settings())
+		sameModel := strings.EqualFold(strings.TrimSpace(route.Model), strings.TrimSpace(chat.Model)) &&
+			strings.EqualFold(strings.TrimSpace(route.Base), strings.TrimSpace(chat.Base))
+		if !rc.Settings().GetBool("runtime.verifierSameModel", false) && sameModel {
+			return gates.Verdict{}, fmt.Errorf("验证器未配置独立轻量模型（当前与主模型相同），跳过同源验证")
+		}
 		if route.Model == "" || route.Base == "" {
 			return gates.Verdict{}, fmt.Errorf("验证器调用失败")
 		}
diff --git a/internal/server/e2e_test.go b/internal/server/e2e_test.go
index 5cae00b..0b84e36 100644
--- a/internal/server/e2e_test.go
+++ b/internal/server/e2e_test.go
@@ -31,9 +31,9 @@ import (
 	"agent-system-backend/internal/profiles"
 	"agent-system-backend/internal/runmgr"
 	"agent-system-backend/internal/runtime"
+	"agent-system-backend/internal/settings"
 	skillseed "agent-system-backend/internal/skills"
 	"agent-system-backend/internal/squadadapter"
-	"agent-system-backend/internal/settings"
 	"agent-system-backend/internal/store"
 	"agent-system-backend/internal/tools"
 	core "agent-system-core"
@@ -929,7 +929,7 @@ func TestSquadOrchestration(t *testing.T) {
 		py, _ := osexec.LookPath("python3")
 		return nativeadp.Config{DefaultModel: "stub", SkillsRoots: []string{skillsRoot},
 			Exec: &nativetools.ExecConfig{Bash: "/bin/bash", Python: py}, Stream: stubNativeStream(
-			"# 交付\n\n这是内核桩后端产出的步骤交付物正文，长度超过五十个字符以满足最小质检要求。")}, true
+				"# 交付\n\n这是内核桩后端产出的步骤交付物正文，长度超过五十个字符以满足最小质检要求。")}, true
 	}
 	defer func() { squadadapter.AgentNativeOptions = savedNative }()
 	h := srv.Handler()
diff --git a/internal/settings/settings.go b/internal/settings/settings.go
index a28b995..050a083 100644
--- a/internal/settings/settings.go
+++ b/internal/settings/settings.go
@@ -96,6 +96,8 @@ func Schema() []core.SettingDesc {
 		{Key: "runtime.verify", Group: "runtime", Label: "终轮结果检查", Type: core.SettingBool, Default: true, Effect: core.EffectNewRun, Description: "终轮确定性检查与独立验证"},
 		{Key: "runtime.llmRetries", Group: "runtime", Label: "模型重试次数", Type: core.SettingNumber, Default: 2, Effect: core.EffectNewRun, Description: "直连模型在还没有任何输出时，遇到超时、限流或服务端错误的额外重试次数。0 = 不重试。已经开始输出后不再重试"},
 		{Key: "runtime.codeOrchestration", Group: "runtime", Label: "代码编排", Type: core.SettingBool, Default: false, Effect: core.EffectNewRun, Description: "允许模型用 run_tool_script 写脚本批量调用工具。默认关闭"},
+		{Key: "runtime.factCheck", Group: "runtime", Label: "终轮事实核查", Type: core.SettingBool, Default: true, Effect: core.EffectNewRun, Description: "交付物里的百分比/日期/大额数字回溯本轮材料，找不到出处的给模型一轮修正（修正过仍失败不拦截，只如实记录）"},
+		{Key: "runtime.verifierSameModel", Group: "runtime", Label: "允许同源验证器", Type: core.SettingBool, Default: false, Effect: core.EffectNewRun, Description: "轻量档未单独配置时，默认跳过与主模型同源的独立验证（硬事实与事实核查仍执行）；显式开启才用主模型做验证"},
 		{Key: "runtime.composeFinal", Group: "runtime", Label: "终轮整理回复", Type: core.SettingBool, Default: false, Effect: core.EffectNewRun, Description: "用过工具的回合结束后，再用主模型重写一段最终回复。默认关闭（裁定 4：转录权威后工作模型已看到完整证据，重写只见过 160 字摘要、实测会虚构章节与承诺未做之事）；异常收尾时另有兜底"},
 		{Key: "runtime.intentSemantic", Group: "runtime", Label: "意图语义裁决（慢通道）", Type: core.SettingBool, Default: false, Effect: core.EffectRestart, Description: "规则歧义时用 embedding 语义相似度裁决。需已配置 embedding 端点。默认关闭（行为与 I1 一致）"},
 		{Key: "runtime.intentGenerator", Group: "runtime", Label: "意图命令生成（慢通道）", Type: core.SettingBool, Default: false, Effect: core.EffectRestart, Description: "规则与语义都未命中时用轻量模型生成意图命令（幻觉 fail-closed 回退）。默认关闭（行为与 I1 一致）"},
diff --git a/internal/squadadapter/adapter_test.go b/internal/squadadapter/adapter_test.go
index e61c06d..852f2e2 100644
--- a/internal/squadadapter/adapter_test.go
+++ b/internal/squadadapter/adapter_test.go
@@ -167,7 +167,6 @@ func containsStr(list []string, want string) bool {
 	return false
 }
 
-
 // Q3 WP2：agent-system 退役后，未注入 AgentNativeOptions 时解析器为空——
 // 步骤后端解析 fail-closed，不再静默走 SpawnSubagent 压缩路径。
 func TestBuildBackendsWithoutNativeOptionsIsEmpty(t *testing.T) {
diff --git a/internal/squadadapter/orchestrate_native_test.go b/internal/squadadapter/orchestrate_native_test.go
index 60906c0..77ba5df 100644
--- a/internal/squadadapter/orchestrate_native_test.go
+++ b/internal/squadadapter/orchestrate_native_test.go
@@ -54,7 +54,6 @@ func TestOrchestrateRunsOnKernelStub(t *testing.T) {
 	}
 }
 
-
 func stubStream(text string) agentllm.StreamFunc {
 	return func(_ context.Context, _ string, msgs []agentllm.Message, _ []agentllm.ToolDecl, _ agentllm.StreamOptions) <-chan agentllm.Event {
 		ch := make(chan agentllm.Event, 8)
diff --git a/internal/tools/exec_backend_test.go b/internal/tools/exec_backend_test.go
index fc25ef4..03d0c85 100644
--- a/internal/tools/exec_backend_test.go
+++ b/internal/tools/exec_backend_test.go
@@ -36,7 +36,7 @@ func TestSandboxArgvShape(t *testing.T) {
 	argv := sandboxArgv(ExecSpec{
 		Bin: "/usr/bin/python3", Args: []string{"/data/s/x.py"},
 		Dir: "/data/s", SkillsDir: "/data/skills",
-		Env:  []string{"PATH=/usr/bin:/bin", "HOME=/tmp/h"},
+		Env: []string{"PATH=/usr/bin:/bin", "HOME=/tmp/h"},
 	})
 	joined := strings.Join(argv, "\x00")
 	musts := []string{
diff --git a/internal/tools/http_test.go b/internal/tools/http_test.go
index 295d883..632685f 100644
--- a/internal/tools/http_test.go
+++ b/internal/tools/http_test.go
@@ -20,9 +20,9 @@ func (dummyCtx) PeerConversations() ([]core.Conversation, error) { return nil, n
 func (dummyCtx) PeerConversationMessages(id string, limit int) ([]core.Message, error) {
 	return nil, nil
 }
-func (dummyCtx) OwnerID() string                       { return "o" }
-func (dummyCtx) WorkspaceRoot() string                 { return "" }
-func (dummyCtx) SettingBool(string, bool) bool         { return false }
+func (dummyCtx) OwnerID() string               { return "o" }
+func (dummyCtx) WorkspaceRoot() string         { return "" }
+func (dummyCtx) SettingBool(string, bool) bool { return false }
 
 // TestHTTPExecutorTemplateAndPath：模板变量注入、结果路径裁剪、错误状态。
 func TestHTTPExecutorTemplateAndPath(t *testing.T) {
diff --git a/internal/tools/search.go b/internal/tools/search.go
index e4b3a97..1315acc 100644
--- a/internal/tools/search.go
+++ b/internal/tools/search.go
@@ -22,7 +22,23 @@ type TavilySettings struct {
 // RegisterSearchTools 注册 tavily_search（完整参数面）与 web_search（别名）。
 // 工具面与 agent-native runtime/tools/search.go 逐字段对齐（名称/参数/输出/失败降级文案）。
 func RegisterSearchTools(r *Registry, provider agenttools.SearchProvider) error {
+	return registerSearch(r, func(ToolExecCtx) agenttools.SearchProvider { return provider })
+}
+
+// RegisterLazySearchTools 注册按调用时取密钥的搜索工具：provider 在每次调用时
+// 重建（设置页/svc 播种改了 tavily 密钥即刻生效，不再需要重启进程——
+// 实测播种后 tavily 工具要重启才注册）。configure 返回 nil 时不静默降级，
+// 返回明确的未配置提示。
+func RegisterLazySearchTools(r *Registry, configure func() agenttools.SearchProvider) error {
+	return registerSearch(r, func(ToolExecCtx) agenttools.SearchProvider { return configure() })
+}
+
+func registerSearch(r *Registry, providerFor func(ToolExecCtx) agenttools.SearchProvider) error {
 	tavily := func(rc ToolExecCtx, args json.RawMessage) (string, error) {
+		provider := providerFor(rc)
+		if provider == nil {
+			return "", fmt.Errorf("搜索未配置：settings.tools.tavily.apiKeys（或对应 secret）为空；配置后立即生效，无需重启")
+		}
 		return execSearch(provider, args)
 	}
 	if err := r.Register(core.ToolDesc{
diff --git a/internal/wfcompile/ab_demo_test.go b/internal/wfcompile/ab_demo_test.go
index 1fbe7bd..081b81b 100644
--- a/internal/wfcompile/ab_demo_test.go
+++ b/internal/wfcompile/ab_demo_test.go
@@ -6,9 +6,9 @@ import (
 	"testing"
 	"time"
 
-	core "agent-system-core"
 	"agent-system-backend/eval"
 	"agent-system-backend/internal/workflow"
+	core "agent-system-core"
 )
 
 // ab_demo_test.go — Q3 WP5 验收：同批任务（含 demo.brief 场景）经两个入口
diff --git a/internal/wfcompile/compiler.go b/internal/wfcompile/compiler.go
index bfc0613..3dfb8ea 100644
--- a/internal/wfcompile/compiler.go
+++ b/internal/wfcompile/compiler.go
@@ -35,20 +35,20 @@ import (
 
 // 宿主 intent 步的意图 ID（目录由本包合成，处理函数在 host.go）。
 const (
-	IntentHitl    = "wf.hitl"
-	IntentTool    = "wf.tool"
+	IntentHitl     = "wf.hitl"
+	IntentTool     = "wf.tool"
 	IntentArtifact = "wf.artifact"
-	IntentCard    = "wf.card"
-	IntentScript  = "wf.script"
-	IntentCritic  = "wf.critic"
+	IntentCard     = "wf.card"
+	IntentScript   = "wf.script"
+	IntentCritic   = "wf.critic"
 )
 
 // CriticPair 记录一个「原步 ↔ 评审步」映射（收口后校验最终判定用）。
 type CriticPair struct {
-	StepID      string // 被评审的原步 ID
-	StepTitle   string
-	EvalID      string      // 评审步 ID（<StepID>_critic）
-	VerdictField string     // 黑板判定字段名
+	StepID       string // 被评审的原步 ID
+	StepTitle    string
+	EvalID       string // 评审步 ID（<StepID>_critic）
+	VerdictField string // 黑板判定字段名
 }
 
 // Compiled 是一次编译产物。
@@ -86,7 +86,7 @@ func Compile(def core.WorkflowDef, vars workflow.Vars, opts CompileOptions, know
 	var steps []squad.StepSpec
 	var agents []squad.RoleSpec
 	agentSeen := map[string]bool{}
-	prevDef := ""  // 上一个 def 步 ID（artifact 空内容的 prev 回退）
+	prevDef := ""   // 上一个 def 步 ID（artifact 空内容的 prev 回退）
 	prevChain := "" // 图链上的前驱（含评审步）
 
 	for i, raw := range def.Steps {
@@ -376,8 +376,8 @@ func buildCatalog() *squad.IntentCatalog {
 		Revision: 1,
 		Markdown: "# 工作流宿主能力\n\n- wf.hitl 人工问答\n- wf.tool 宿主工具\n- wf.artifact 工件落库\n- wf.card 收件箱卡\n- wf.script 脚本宿主\n- wf.critic 步骤验收\n",
 		Specs: map[string]squad.IntentSpecView{
-			IntentHitl: spec(IntentHitl, "人工问答步", req("question"), opt("step_id"), opt("kind"), opt("choices"), opt("allow_free_input"), opt("free_input_hint")),
-			IntentTool: spec(IntentTool, "宿主工具步", req("tool"), opt("step_id"), opt("args")),
+			IntentHitl:     spec(IntentHitl, "人工问答步", req("question"), opt("step_id"), opt("kind"), opt("choices"), opt("allow_free_input"), opt("free_input_hint")),
+			IntentTool:     spec(IntentTool, "宿主工具步", req("tool"), opt("step_id"), opt("args")),
 			IntentArtifact: spec(IntentArtifact, "工件步", req("name"), opt("step_id"), opt("content"), opt("content_ref")),
 			IntentCard:     spec(IntentCard, "收件箱卡步", req("template"), opt("step_id")),
 			IntentScript:   spec(IntentScript, "脚本步", req("path"), opt("step_id"), opt("model")),