diff --git a/server/api/src/__tests__/agents/graphs/ingest/formatResearchForIngest.test.ts b/server/api/src/__tests__/agents/graphs/ingest/formatResearchForIngest.test.ts
index fcb0afea..1bc7dd26 100644
--- a/server/api/src/__tests__/agents/graphs/ingest/formatResearchForIngest.test.ts
+++ b/server/api/src/__tests__/agents/graphs/ingest/formatResearchForIngest.test.ts
@@ -24,7 +24,7 @@ function baseState(): IngestPlannerStateType {
         iteration: 1,
         queries: [],
         sources: [],
-        evaluation: { score: 0.9, rationale: "ok", missingAspects: [] },
+        evaluation: { score: 0.9, sufficient: true, rationale: "ok", missingAspects: [] },
         createdAt: "2026-01-01T00:00:00.000Z",
       },
     ],
diff --git a/server/api/src/__tests__/agents/graphs/ingest/ingestPlannerGraph.test.ts b/server/api/src/__tests__/agents/graphs/ingest/ingestPlannerGraph.test.ts
index a4f0ccaa..30a1bb17 100644
--- a/server/api/src/__tests__/agents/graphs/ingest/ingestPlannerGraph.test.ts
+++ b/server/api/src/__tests__/agents/graphs/ingest/ingestPlannerGraph.test.ts
@@ -102,7 +102,7 @@ function defaultMocks() {
   wikiSearch.mockImplementation(async () => ({ pendingSources: [] }));
   fetchArticles.mockImplementation(async () => ({ pendingSources: [] }));
   evaluateSufficiency.mockImplementation(async (state: { iteration: number }) => ({
-    lastEvaluation: { score: 0.9, rationale: "ok", missingAspects: [] },
+    lastEvaluation: { score: 0.9, sufficient: true, rationale: "ok", missingAspects: [] },
     iteration: state.iteration + 1,
     phase: "research:evaluated",
   }));
diff --git a/server/api/src/__tests__/agents/graphs/wikiCompose/nodes/humanReviewBrief.test.ts b/server/api/src/__tests__/agents/graphs/wikiCompose/nodes/humanReviewBrief.test.ts
index 9c0c8eda..70464f2e 100644
--- a/server/api/src/__tests__/agents/graphs/wikiCompose/nodes/humanReviewBrief.test.ts
+++ b/server/api/src/__tests__/agents/graphs/wikiCompose/nodes/humanReviewBrief.test.ts
@@ -74,7 +74,6 @@ describe("humanReviewBrief", () => {
         },
       ],
       appendToExisting: true,
-      researchMaxIterations: 4,
     });
 
     const update = await humanReviewBrief(state({}), { configurable: {} } as never);
@@ -97,7 +96,6 @@ describe("humanReviewBrief", () => {
     ]);
     expect(brief.summary).toContain("Target audience?");
     expect(brief.summary).toContain("selected=opt-a");
-    expect(update.maxIterations).toBe(4);
   });
 
   it("accepts an empty answers array (explicit Brief skip)", async () => {
@@ -111,13 +109,6 @@ describe("humanReviewBrief", () => {
     expect(brief.answers).toEqual([]);
     expect(brief.summary).toBe("(no brief provided)");
     expect(brief.appendToExisting).toBe(false);
-    expect(update.maxIterations).toBeUndefined();
-  });
-
-  it("rejects researchMaxIterations outside 1..5", async () => {
-    interrupt.mockReturnValueOnce({ answers: [], researchMaxIterations: 9 });
-
-    await expect(humanReviewBrief(state({}), { configurable: {} } as never)).rejects.toThrow();
   });
 
   it("rejects answers missing questionId", async () => {
diff --git a/server/api/src/__tests__/agents/graphs/wikiCompose/wikiComposeGraph.test.ts b/server/api/src/__tests__/agents/graphs/wikiCompose/wikiComposeGraph.test.ts
index 46ba64f9..d5c259ed 100644
--- a/server/api/src/__tests__/agents/graphs/wikiCompose/wikiComposeGraph.test.ts
+++ b/server/api/src/__tests__/agents/graphs/wikiCompose/wikiComposeGraph.test.ts
@@ -124,7 +124,7 @@ function defaultMocks() {
   wikiSearch.mockImplementation(async () => ({ pendingSources: [] }));
   fetchArticles.mockImplementation(async () => ({ pendingSources: [] }));
   evaluateSufficiency.mockImplementation(async (state: { iteration: number }) => ({
-    lastEvaluation: { score: 0.9, rationale: "ok", missingAspects: [] },
+    lastEvaluation: { score: 0.9, sufficient: true, rationale: "ok", missingAspects: [] },
     iteration: state.iteration + 1,
     phase: "research:evaluated",
   }));
diff --git a/server/api/src/__tests__/agents/subgraphs/research/constants.test.ts b/server/api/src/__tests__/agents/subgraphs/research/constants.test.ts
new file mode 100644
index 00000000..6631d2f6
--- /dev/null
+++ b/server/api/src/__tests__/agents/subgraphs/research/constants.test.ts
@@ -0,0 +1,38 @@
+/**
+ * Research loop iteration cap resolution tests.
+ */
+import { describe, expect, it } from "vitest";
+import {
+  INGEST_RESEARCH_GRAPH_ID,
+  RESEARCH_SAFETY_MAX_ITERATIONS,
+  clampIngestMaxIterations,
+  resolveResearchMaxIterations,
+} from "../../../../agents/subgraphs/research/constants.js";
+import { WIKI_COMPOSE_GRAPH_ID } from "../../../../agents/graphs/wikiCompose/index.js";
+
+describe("clampIngestMaxIterations", () => {
+  it("clamps ingest caps to 1..5 with default 3", () => {
+    expect(clampIngestMaxIterations(undefined)).toBe(3);
+    expect(clampIngestMaxIterations(99)).toBe(5);
+    expect(clampIngestMaxIterations(4)).toBe(4);
+  });
+});
+
+describe("resolveResearchMaxIterations", () => {
+  it("honours ingest graph caps from state", () => {
+    expect(resolveResearchMaxIterations(INGEST_RESEARCH_GRAPH_ID, 4)).toBe(4);
+    expect(resolveResearchMaxIterations(INGEST_RESEARCH_GRAPH_ID, 99)).toBe(5);
+  });
+
+  it("uses the safety cap for Wiki Compose graphs regardless of legacy state", () => {
+    expect(resolveResearchMaxIterations(WIKI_COMPOSE_GRAPH_ID, undefined)).toBe(
+      RESEARCH_SAFETY_MAX_ITERATIONS,
+    );
+    expect(resolveResearchMaxIterations(WIKI_COMPOSE_GRAPH_ID, 3)).toBe(
+      RESEARCH_SAFETY_MAX_ITERATIONS,
+    );
+    expect(resolveResearchMaxIterations("wiki-compose-research", 3)).toBe(
+      RESEARCH_SAFETY_MAX_ITERATIONS,
+    );
+  });
+});
diff --git a/server/api/src/__tests__/agents/subgraphs/research/nodes/compileBatch.test.ts b/server/api/src/__tests__/agents/subgraphs/research/nodes/compileBatch.test.ts
index 005c3cb9..9ebb0986 100644
--- a/server/api/src/__tests__/agents/subgraphs/research/nodes/compileBatch.test.ts
+++ b/server/api/src/__tests__/agents/subgraphs/research/nodes/compileBatch.test.ts
@@ -1,16 +1,12 @@
 /**
  * `compileBatch` unit tests. Pure projection node; no LLM. We verify:
- * - `exitReason` = "score_threshold" when score >= 0.75.
- * - `exitReason` = "max_iterations" otherwise.
+ * - `exitReason` = "score_threshold" when evaluation is sufficient.
+ * - `exitReason` = "safety_cap" for Wiki Compose iteration cap hits.
+ * - `exitReason` = "max_iterations" for ingest iteration cap hits.
  * - Batch fields are populated from state.
- * - `dispatchCustomEvent` is called via the runnable config.
  */
 import { describe, expect, it, vi } from "vitest";
 
-// The dispatch helper requires a proper LangChain callback manager which we
-// don't set up here (`compileBatch` is a pure projection). Stub it so the
-// node can dispatch into a no-op without a real callback runtime.
-// dispatch ヘルパは callback manager 必須なので test では no-op に差し替える。
 const { dispatchResearchBatch } = vi.hoisted(() => ({
   dispatchResearchBatch: vi.fn(async () => undefined),
 }));
@@ -20,9 +16,15 @@ vi.mock("../../../../../agents/subgraphs/research/nodes/shared/dispatchSseCustom
   dispatchResearchIteration: vi.fn(),
 }));
 
-import { compileBatch } from "../../../../../agents/subgraphs/research/nodes/compileBatch.js";
+import {
+  compileBatch,
+  resolveResearchExitReason,
+} from "../../../../../agents/subgraphs/research/nodes/compileBatch.js";
+import { GRAPH_CONTEXT_CONFIG_KEY } from "../../../../../agents/core/types/graphContext.js";
+import type { GraphContext } from "../../../../../agents/core/types/graphContext.js";
 import type { ResearchLoopStateType } from "../../../../../agents/subgraphs/research/state.js";
 import type { ResearchBatch } from "../../../../../agents/subgraphs/research/types.js";
+import type { Database } from "../../../../../types/index.js";
 
 function state(overrides: Partial<ResearchLoopStateType>): ResearchLoopStateType {
   return {
@@ -47,18 +49,75 @@ function state(overrides: Partial<ResearchLoopStateType>): ResearchLoopStateType
   };
 }
 
+function configForGraph(graphId: string) {
+  const ctx: GraphContext = {
+    threadId: "t",
+    sessionId: "t",
+    userId: "user-1",
+    pageId: "page-1",
+    graphId,
+    backend: "zedi_managed",
+    tier: "free",
+    db: {} as Database,
+    feature: "wiki_compose:research",
+    userEmail: null,
+    contentLocale: "ja",
+  };
+  return { configurable: { [GRAPH_CONTEXT_CONFIG_KEY]: ctx } };
+}
+
+describe("resolveResearchExitReason", () => {
+  it("returns score_threshold when sufficient is true", () => {
+    expect(
+      resolveResearchExitReason(
+        state({
+          lastEvaluation: { score: 0.2, sufficient: true, rationale: "ok", missingAspects: [] },
+        }),
+        "wiki-compose",
+      ),
+    ).toBe("score_threshold");
+  });
+
+  it("returns safety_cap for Wiki Compose iteration cap hits", () => {
+    expect(
+      resolveResearchExitReason(
+        state({
+          lastEvaluation: {
+            score: 0.2,
+            sufficient: false,
+            rationale: "weak",
+            missingAspects: ["x"],
+          },
+        }),
+        "wiki-compose",
+      ),
+    ).toBe("safety_cap");
+  });
+
+  it("returns max_iterations for ingest iteration cap hits", () => {
+    expect(
+      resolveResearchExitReason(
+        state({
+          lastEvaluation: {
+            score: 0.2,
+            sufficient: false,
+            rationale: "weak",
+            missingAspects: ["x"],
+          },
+        }),
+        "ingest-planner",
+      ),
+    ).toBe("max_iterations");
+  });
+});
+
 describe("compileBatch", () => {
-  it("uses score_threshold when last score >= 0.75", async () => {
-    const dispatcher = vi.fn();
-    const config = {
-      configurable: { callbacks: undefined },
-      callbacks: { handlers: [], inheritableHandlers: [], dispatchCustomEvent: dispatcher },
-    };
+  it("uses score_threshold when last evaluation is sufficient", async () => {
     const update = await compileBatch(
-      state({ lastEvaluation: { score: 0.85, rationale: "ok", missingAspects: [] } }),
-      // Loose config type — node only reads callback runtime, which LangGraph
-      // wires through the surrounding `streamEvents` / `invoke` call.
-      config as never,
+      state({
+        lastEvaluation: { score: 0.85, sufficient: true, rationale: "ok", missingAspects: [] },
+      }),
+      configForGraph("wiki-compose") as never,
     );
     expect(update.exitReason).toBe("score_threshold");
     const batches = update.batches as ResearchBatch[] | undefined;
@@ -67,19 +126,32 @@ describe("compileBatch", () => {
     expect(batches?.[0]?.iteration).toBe(2);
   });
 
-  it("uses max_iterations when no eval or score below threshold", async () => {
+  it("uses safety_cap for Wiki Compose when evaluation is insufficient", async () => {
     const update = await compileBatch(
-      state({ lastEvaluation: { score: 0.5, rationale: "weak", missingAspects: ["x"] } }),
-      { configurable: {} } as never,
+      state({
+        lastEvaluation: { score: 0.5, sufficient: false, rationale: "weak", missingAspects: ["x"] },
+      }),
+      configForGraph("wiki-compose-research") as never,
+    );
+    expect(update.exitReason).toBe("safety_cap");
+  });
+
+  it("uses max_iterations for ingest when evaluation is insufficient", async () => {
+    const update = await compileBatch(
+      state({
+        lastEvaluation: { score: 0.5, sufficient: false, rationale: "weak", missingAspects: ["x"] },
+      }),
+      configForGraph("ingest-planner") as never,
     );
     expect(update.exitReason).toBe("max_iterations");
   });
 
   it("handles null evaluation gracefully", async () => {
-    const update = await compileBatch(state({ lastEvaluation: null }), {
-      configurable: {},
-    } as never);
-    expect(update.exitReason).toBe("max_iterations");
+    const update = await compileBatch(
+      state({ lastEvaluation: null }),
+      configForGraph("wiki-compose") as never,
+    );
+    expect(update.exitReason).toBe("safety_cap");
     const batches = update.batches as ResearchBatch[] | undefined;
     expect(batches?.[0]?.evaluation).toBeNull();
   });
diff --git a/server/api/src/__tests__/agents/subgraphs/research/nodes/planQueries.test.ts b/server/api/src/__tests__/agents/subgraphs/research/nodes/planQueries.test.ts
index afaa639a..071f45e8 100644
--- a/server/api/src/__tests__/agents/subgraphs/research/nodes/planQueries.test.ts
+++ b/server/api/src/__tests__/agents/subgraphs/research/nodes/planQueries.test.ts
@@ -32,6 +32,11 @@ vi.mock("../../../../../agents/subgraphs/research/nodes/shared/dispatchSseCustom
 }));
 
 import { planQueries } from "../../../../../agents/subgraphs/research/nodes/planQueries.js";
+import {
+  INGEST_RESEARCH_GRAPH_ID,
+  RESEARCH_SAFETY_MAX_ITERATIONS,
+} from "../../../../../agents/subgraphs/research/constants.js";
+import { WIKI_COMPOSE_GRAPH_ID } from "../../../../../agents/graphs/wikiCompose/index.js";
 import { GRAPH_CONTEXT_CONFIG_KEY } from "../../../../../agents/core/types/graphContext.js";
 import type { GraphContext } from "../../../../../agents/core/types/graphContext.js";
 import type { Database } from "../../../../../types/index.js";
@@ -93,9 +98,25 @@ afterEach(() => {
 describe("planQueries — additional research detection", () => {
   const config = { configurable: { [GRAPH_CONTEXT_CONFIG_KEY]: fakeContext() } };
 
-  it("clamps maxIterations to 1..5 (default 3)", async () => {
+  it("uses the safety cap for Wiki Compose even when legacy state has maxIterations: 3", async () => {
+    const ctx = fakeContext();
+    ctx.graphId = WIKI_COMPOSE_GRAPH_ID;
+    const wikiConfig = { configurable: { [GRAPH_CONTEXT_CONFIG_KEY]: ctx } };
+    const update = await planQueries(state({ maxIterations: 3 }), wikiConfig as never);
+    expect(update.maxIterations).toBe(RESEARCH_SAFETY_MAX_ITERATIONS);
+  });
+
+  it("uses the safety cap when graph is standalone research", async () => {
     const update = await planQueries(state({ maxIterations: 99 }), config as never);
-    expect(update.maxIterations).toBe(5);
+    expect(update.maxIterations).toBe(RESEARCH_SAFETY_MAX_ITERATIONS);
+  });
+
+  it("honours ingest graph caps from state", async () => {
+    const ctx = fakeContext();
+    ctx.graphId = INGEST_RESEARCH_GRAPH_ID;
+    const ingestConfig = { configurable: { [GRAPH_CONTEXT_CONFIG_KEY]: ctx } };
+    const update = await planQueries(state({ maxIterations: 4 }), ingestConfig as never);
+    expect(update.maxIterations).toBe(4);
   });
 
   it("consumes state.additionalRequest and seeds carried-over sources", async () => {
diff --git a/server/api/src/__tests__/agents/subgraphs/research/researchGraph.conditional.test.ts b/server/api/src/__tests__/agents/subgraphs/research/researchGraph.conditional.test.ts
index ac8fa778..e9320a86 100644
--- a/server/api/src/__tests__/agents/subgraphs/research/researchGraph.conditional.test.ts
+++ b/server/api/src/__tests__/agents/subgraphs/research/researchGraph.conditional.test.ts
@@ -37,7 +37,7 @@ describe("shouldRefine", () => {
         state({
           iteration: 1,
           maxIterations: 5,
-          lastEvaluation: { score: 0.75, rationale: "ok", missingAspects: [] },
+          lastEvaluation: { score: 0.75, sufficient: true, rationale: "ok", missingAspects: [] },
         }),
       ),
     ).toBe("compile");
@@ -49,7 +49,7 @@ describe("shouldRefine", () => {
         state({
           iteration: 1,
           maxIterations: 5,
-          lastEvaluation: { score: 0.95, rationale: "great", missingAspects: [] },
+          lastEvaluation: { score: 0.95, sufficient: true, rationale: "great", missingAspects: [] },
         }),
       ),
     ).toBe("compile");
@@ -61,7 +61,12 @@ describe("shouldRefine", () => {
         state({
           iteration: 1,
           maxIterations: 3,
-          lastEvaluation: { score: 0.5, rationale: "weak", missingAspects: ["x"] },
+          lastEvaluation: {
+            score: 0.5,
+            sufficient: false,
+            rationale: "weak",
+            missingAspects: ["x"],
+          },
         }),
       ),
     ).toBe("refine");
@@ -73,7 +78,12 @@ describe("shouldRefine", () => {
         state({
           iteration: 3,
           maxIterations: 3,
-          lastEvaluation: { score: 0.4, rationale: "weak", missingAspects: ["x", "y"] },
+          lastEvaluation: {
+            score: 0.4,
+            sufficient: false,
+            rationale: "weak",
+            missingAspects: ["x", "y"],
+          },
         }),
       ),
     ).toBe("compile");
@@ -83,6 +93,18 @@ describe("shouldRefine", () => {
     expect(shouldRefine(state({ iteration: 4, maxIterations: 3 }))).toBe("compile");
   });
 
+  it("compiles when sufficient is true even if score is below threshold", () => {
+    expect(
+      shouldRefine(
+        state({
+          iteration: 1,
+          maxIterations: 5,
+          lastEvaluation: { score: 0.4, sufficient: true, rationale: "ok", missingAspects: [] },
+        }),
+      ),
+    ).toBe("compile");
+  });
+
   it("refines when there's no evaluation yet and iterations remain", () => {
     // Defensive: if evaluate_sufficiency hasn't run, treat as "not enough yet".
     // evaluation 未走の保険 — まだ充足してないとみなす。
diff --git a/server/api/src/__tests__/agents/subgraphs/research/researchGraph.interrupt.test.ts b/server/api/src/__tests__/agents/subgraphs/research/researchGraph.interrupt.test.ts
index c0ffdb53..c72f5abf 100644
--- a/server/api/src/__tests__/agents/subgraphs/research/researchGraph.interrupt.test.ts
+++ b/server/api/src/__tests__/agents/subgraphs/research/researchGraph.interrupt.test.ts
@@ -101,7 +101,7 @@ describe("researchLoopSubgraph — interrupt at human_review_research", () => {
     wikiSearch.mockImplementation(async () => ({ pendingSources: [] }));
     fetchArticles.mockImplementation(async () => ({ pendingSources: [] }));
     evaluateSufficiency.mockImplementation(async (state, _c) => ({
-      lastEvaluation: { score: 0.9, rationale: "ok", missingAspects: [] },
+      lastEvaluation: { score: 0.9, sufficient: true, rationale: "ok", missingAspects: [] },
       iteration: state.iteration + 1,
       phase: "research:evaluated",
     }));
diff --git a/server/api/src/__tests__/agents/subgraphs/research/researchGraph.loop.test.ts b/server/api/src/__tests__/agents/subgraphs/research/researchGraph.loop.test.ts
index 4b1fcf44..873aeeb1 100644
--- a/server/api/src/__tests__/agents/subgraphs/research/researchGraph.loop.test.ts
+++ b/server/api/src/__tests__/agents/subgraphs/research/researchGraph.loop.test.ts
@@ -124,7 +124,7 @@ describe("researchLoopSubgraph — autonomous loop", () => {
     evaluateSufficiency.mockImplementation(async (state, _config) => {
       evaluatedTimes += 1;
       return {
-        lastEvaluation: { score: 0.1, rationale: "weak", missingAspects: ["x"] },
+        lastEvaluation: { score: 0.1, sufficient: false, rationale: "weak", missingAspects: ["x"] },
         iteration: state.iteration + 1,
         phase: "research:evaluated",
       };
@@ -194,7 +194,7 @@ describe("researchLoopSubgraph — autonomous loop", () => {
     wikiSearch.mockImplementation(async () => ({ pendingSources: [] }));
     fetchArticles.mockImplementation(async () => ({ pendingSources: [] }));
     evaluateSufficiency.mockImplementation(async (state, _c) => ({
-      lastEvaluation: { score: 0.9, rationale: "great", missingAspects: [] },
+      lastEvaluation: { score: 0.9, sufficient: true, rationale: "great", missingAspects: [] },
       iteration: state.iteration + 1,
       phase: "research:evaluated",
     }));
diff --git a/server/api/src/__tests__/agents/subgraphs/research/researchGraph.modelGuard.test.ts b/server/api/src/__tests__/agents/subgraphs/research/researchGraph.modelGuard.test.ts
index edd3a053..ecc5b0c9 100644
--- a/server/api/src/__tests__/agents/subgraphs/research/researchGraph.modelGuard.test.ts
+++ b/server/api/src/__tests__/agents/subgraphs/research/researchGraph.modelGuard.test.ts
@@ -138,6 +138,7 @@ describe("researchLoopSubgraph — all LLM calls go through ZediChatModel", () =
         const score = evaluateCall >= 2 ? 0.9 : 0.1;
         return fakeModel(async () => ({
           score,
+          sufficient: score >= 0.75,
           rationale: "auto",
           missingAspects: score < 0.75 ? ["x"] : [],
         }));
diff --git a/server/api/src/__tests__/agents/subgraphs/research/researchGraph.resume.test.ts b/server/api/src/__tests__/agents/subgraphs/research/researchGraph.resume.test.ts
index 0385ff89..4f328780 100644
--- a/server/api/src/__tests__/agents/subgraphs/research/researchGraph.resume.test.ts
+++ b/server/api/src/__tests__/agents/subgraphs/research/researchGraph.resume.test.ts
@@ -105,7 +105,7 @@ describe("researchLoopSubgraph — resume projects approvedResearch", () => {
     wikiSearch.mockImplementation(async () => ({ pendingSources: pending.slice(2) }));
     fetchArticles.mockImplementation(async () => ({ pendingSources: [] }));
     evaluateSufficiency.mockImplementation(async (state, _c) => ({
-      lastEvaluation: { score: 0.9, rationale: "ok", missingAspects: [] },
+      lastEvaluation: { score: 0.9, sufficient: true, rationale: "ok", missingAspects: [] },
       iteration: state.iteration + 1,
       phase: "research:evaluated",
     }));
@@ -189,7 +189,7 @@ describe("researchLoopSubgraph — resume projects approvedResearch", () => {
     wikiSearch.mockImplementation(async () => ({ pendingSources: [] }));
     fetchArticles.mockImplementation(async () => ({ pendingSources: [] }));
     evaluateSufficiency.mockImplementation(async (state, _c) => ({
-      lastEvaluation: { score: 0.9, rationale: "ok", missingAspects: [] },
+      lastEvaluation: { score: 0.9, sufficient: true, rationale: "ok", missingAspects: [] },
       iteration: state.iteration + 1,
       phase: "research:evaluated",
     }));
diff --git a/server/api/src/agents/core/types/sseEvents.ts b/server/api/src/agents/core/types/sseEvents.ts
index f2d83d92..56df62ce 100644
--- a/server/api/src/agents/core/types/sseEvents.ts
+++ b/server/api/src/agents/core/types/sseEvents.ts
@@ -164,7 +164,7 @@ export interface SseResearchBatchEvent {
   /** Last evaluation score (null only if compile fired before any evaluate). */
   score: number | null;
   /** Reason the loop exited. */
-  exitReason: "score_threshold" | "max_iterations";
+  exitReason: "score_threshold" | "max_iterations" | "safety_cap";
 }
 
 /**
diff --git a/server/api/src/agents/graphs/ingest/nodes/prepareIngest.ts b/server/api/src/agents/graphs/ingest/nodes/prepareIngest.ts
index 5ca4579e..076f8766 100644
--- a/server/api/src/agents/graphs/ingest/nodes/prepareIngest.ts
+++ b/server/api/src/agents/graphs/ingest/nodes/prepareIngest.ts
@@ -7,14 +7,9 @@
 import { HumanMessage } from "@langchain/core/messages";
 import type { LangGraphRunnableConfig } from "@langchain/langgraph";
 import { getGraphContext } from "../../../subgraphs/research/nodes/shared/getGraphContext.js";
+import { clampIngestMaxIterations } from "../../../subgraphs/research/constants.js";
 import type { IngestPlannerStateType, IngestPlannerStateUpdate } from "../state.js";
 
-function clampMaxIterations(raw: number): number {
-  if (!Number.isFinite(raw)) return 3;
-  const truncated = Math.trunc(raw);
-  return Math.min(Math.max(truncated, 1), 5);
-}
-
 /**
  * Project graph run input into ingest + research seed state.
  *
@@ -34,7 +29,7 @@ export async function prepareIngest(
 
   const candidates = state.candidates;
   const userSchema = state.userSchema;
-  const maxIterations = clampMaxIterations(state.maxIterations);
+  const maxIterations = clampIngestMaxIterations(state.maxIterations);
 
   for (const [i, c] of candidates.entries()) {
     if (!c?.id?.trim() || typeof c.title !== "string" || !c.title.trim()) {
diff --git a/server/api/src/agents/graphs/wikiCompose/nodes/humanReviewBrief.ts b/server/api/src/agents/graphs/wikiCompose/nodes/humanReviewBrief.ts
index 831ea038..941726c3 100644
--- a/server/api/src/agents/graphs/wikiCompose/nodes/humanReviewBrief.ts
+++ b/server/api/src/agents/graphs/wikiCompose/nodes/humanReviewBrief.ts
@@ -4,13 +4,10 @@
  * Brief 質問群を `interrupt(value)` でユーザーに渡し、`PATCH .../resume` の
  * 結果を `briefResumeSchema` で検証して `brief` を state に確定する。
  * 既存本文ありで「追記」を選んだ場合は `appendToExisting=true` が立ち、Draft
- * フェーズがそれを読んで挙動を切り替える。`researchMaxIterations` (1..5) が
- * 指定されていれば、後段の Research subgraph に渡るようミラーする。
+ * フェーズがそれを読んで挙動を切り替える。
  *
  * Halts the graph at the Brief interrupt and projects the user's answers into
- * `state.brief`. The resume payload's `researchMaxIterations` (when present)
- * is mirrored to `state.researchMaxIterations` so the research subgraph node
- * picks it up via its own state slot when invoked.
+ * `state.brief`.
  */
 import type { LangGraphRunnableConfig } from "@langchain/langgraph";
 import { interrupt } from "@langchain/langgraph";
@@ -127,11 +124,5 @@ export async function humanReviewBrief(
     brief,
     phase: "brief:completed",
   };
-  if (parsed.researchMaxIterations !== undefined) {
-    // Mirror onto the canonical research subgraph channel name so the
-    // composed research node picks it up via shared state.
-    // research subgraph と共有する `maxIterations` チャネルに反映する。
-    update.maxIterations = parsed.researchMaxIterations;
-  }
   return update;
 }
diff --git a/server/api/src/agents/graphs/wikiCompose/resumeSchemas.ts b/server/api/src/agents/graphs/wikiCompose/resumeSchemas.ts
index dec5f9d4..8fa16fa0 100644
--- a/server/api/src/agents/graphs/wikiCompose/resumeSchemas.ts
+++ b/server/api/src/agents/graphs/wikiCompose/resumeSchemas.ts
@@ -16,13 +16,11 @@ import { z } from "zod";
  *
  * - `answers` — 必須。空配列でも可（Brief をスキップしたケース）。
  * - `appendToExisting` — 本文ありページで「追記」を選んだ場合 true。
- * - `researchMaxIterations` — Brief 内で 1..5 にユーザーが調整した場合のみ。
  *
  * Validates the resume payload at the Brief interrupt. `answers` is required
  * even when empty (the user may explicitly skip Brief by submitting an empty
  * array). Default for `appendToExisting` is `false` (replace-mode is the
- * historical Wiki Compose behaviour); `researchMaxIterations` is clamped to
- * 1..5 by the schema so the graph never sees an out-of-range value.
+ * historical Wiki Compose behaviour).
  */
 export const briefResumeSchema = z.object({
   answers: z
@@ -35,7 +33,6 @@ export const briefResumeSchema = z.object({
     )
     .default([]),
   appendToExisting: z.boolean().optional().default(false),
-  researchMaxIterations: z.number().int().min(1).max(5).optional(),
 });
 
 export type BriefResumeParsed = z.infer<typeof briefResumeSchema>;
diff --git a/server/api/src/agents/graphs/wikiCompose/state.ts b/server/api/src/agents/graphs/wikiCompose/state.ts
index 8960d4cd..1d84fea6 100644
--- a/server/api/src/agents/graphs/wikiCompose/state.ts
+++ b/server/api/src/agents/graphs/wikiCompose/state.ts
@@ -19,6 +19,7 @@
  */
 import { Annotation } from "@langchain/langgraph";
 import { BaseState } from "../../core/state/baseState.js";
+import { RESEARCH_SAFETY_MAX_ITERATIONS } from "../../subgraphs/research/constants.js";
 import type {
   AdditionalResearchRequest,
   Evaluation,
@@ -155,10 +156,10 @@ export const WikiComposeState = Annotation.Root({
     reducer: (_prev, next) => next,
     default: () => 0,
   }),
-  /** ループ上限（Brief で 1..5 にユーザー設定可、デフォルト 3）。 */
+  /** ループ上限（自律調査の安全 cap。ingest 連携時のみ 1..5 の明示 cap あり）。 */
   maxIterations: Annotation<number>({
     reducer: (prev, next) => next ?? prev,
-    default: () => 3,
+    default: () => RESEARCH_SAFETY_MAX_ITERATIONS,
   }),
   /** Research subgraph 内の直近クエリ。 */
   queries: Annotation<PlannedQuery[]>({
diff --git a/server/api/src/agents/graphs/wikiCompose/types.ts b/server/api/src/agents/graphs/wikiCompose/types.ts
index a80c0114..cf39622b 100644
--- a/server/api/src/agents/graphs/wikiCompose/types.ts
+++ b/server/api/src/agents/graphs/wikiCompose/types.ts
@@ -272,8 +272,6 @@ export interface BriefResumeInput {
   answers: BriefAnswer[];
   /** True when the user chose "append to existing body" (U2). */
   appendToExisting?: boolean;
-  /** Optional override for the research loop's max iterations (1..5). */
-  researchMaxIterations?: number;
 }
 
 /** Resume payload for the outline interrupt. */
diff --git a/server/api/src/agents/runner/sseMapper.ts b/server/api/src/agents/runner/sseMapper.ts
index 3855f972..60d62a0c 100644
--- a/server/api/src/agents/runner/sseMapper.ts
+++ b/server/api/src/agents/runner/sseMapper.ts
@@ -305,7 +305,9 @@ function mapResearchBatch(data: Record<string, unknown>): SseResearchBatchEvent[
   const sourceCount = typeof data.sourceCount === "number" ? data.sourceCount : null;
   const score = data.score === null || typeof data.score === "number" ? data.score : null;
   const exitReason =
-    data.exitReason === "score_threshold" || data.exitReason === "max_iterations"
+    data.exitReason === "score_threshold" ||
+    data.exitReason === "max_iterations" ||
+    data.exitReason === "safety_cap"
       ? data.exitReason
       : null;
   if (batchId === null || iteration === null || sourceCount === null || exitReason === null) {
diff --git a/server/api/src/agents/subgraphs/research/constants.ts b/server/api/src/agents/subgraphs/research/constants.ts
new file mode 100644
index 00000000..f7467885
--- /dev/null
+++ b/server/api/src/agents/subgraphs/research/constants.ts
@@ -0,0 +1,68 @@
+/**
+ * Research loop limits and shared thresholds for Wiki Compose / ingest graphs.
+ * Wiki Compose / ingest グラフ共通の調査ループ定数。
+ */
+
+/**
+ * Hard safety cap for autonomous Wiki Compose research loops. The evaluator
+ * LLM decides when sources are sufficient; this constant prevents runaway loops.
+ *
+ * 自律調査の安全上限。充足度評価 LLM が十分と判断するまでループし、
+ * 無限ループ防止のためだけにこの上限を使う。
+ */
+export const RESEARCH_SAFETY_MAX_ITERATIONS = 10;
+
+/**
+ * Score at or above which research is treated as sufficient when the evaluator
+ * does not set {@link Evaluation.sufficient} explicitly.
+ *
+ * 評価 LLM が `sufficient` を返さない場合のフォールバック閾値。
+ */
+export const RESEARCH_SUFFICIENCY_SCORE_THRESHOLD = 0.75;
+
+/**
+ * Ingest planner graph id. Keep in sync with
+ * {@link INGEST_PLANNER_GRAPH_ID} in `graphs/ingest/ingestPlannerGraph.ts`.
+ *
+ * ingest グラフ ID。ingest 側の定数と文字列一致を維持する。
+ */
+export const INGEST_RESEARCH_GRAPH_ID = "ingest-planner" as const;
+
+/** Ingest API explicit iteration cap range (1..5). / ingest 明示 cap の範囲。 */
+export const INGEST_EXPLICIT_MAX_ITERATIONS_MIN = 1;
+export const INGEST_EXPLICIT_MAX_ITERATIONS_MAX = 5;
+
+/**
+ * Clamp ingest `maxIterations` input to 1..5 (default 3 when invalid).
+ *
+ * ingest の `maxIterations` を 1..5 にクランプする。
+ *
+ * @param raw Value from ingest graph state / run input.
+ */
+export function clampIngestMaxIterations(raw: unknown): number {
+  if (typeof raw !== "number" || !Number.isFinite(raw)) return 3;
+  const truncated = Math.trunc(raw);
+  return Math.min(
+    Math.max(truncated, INGEST_EXPLICIT_MAX_ITERATIONS_MIN),
+    INGEST_EXPLICIT_MAX_ITERATIONS_MAX,
+  );
+}
+
+/**
+ * Resolve the iteration cap for the research loop from the owning graph id.
+ *
+ * - ingest planner → clamped caller cap (1..5)
+ * - Wiki Compose / standalone research → {@link RESEARCH_SAFETY_MAX_ITERATIONS}
+ *
+ * Legacy checkpoints with `maxIterations: 3` on Wiki Compose graphs are ignored;
+ * only ingest uses state-provided caps.
+ *
+ * @param graphId Owning graph id from {@link GraphContext}.
+ * @param stateMaxIterations Current `state.maxIterations` (ingest only).
+ */
+export function resolveResearchMaxIterations(graphId: string, stateMaxIterations: unknown): number {
+  if (graphId === INGEST_RESEARCH_GRAPH_ID) {
+    return clampIngestMaxIterations(stateMaxIterations);
+  }
+  return RESEARCH_SAFETY_MAX_ITERATIONS;
+}
diff --git a/server/api/src/agents/subgraphs/research/nodes/compileBatch.ts b/server/api/src/agents/subgraphs/research/nodes/compileBatch.ts
index b9561dfc..b952d428 100644
--- a/server/api/src/agents/subgraphs/research/nodes/compileBatch.ts
+++ b/server/api/src/agents/subgraphs/research/nodes/compileBatch.ts
@@ -8,9 +8,26 @@
  */
 import type { LangGraphRunnableConfig } from "@langchain/langgraph";
 import { randomUUID } from "node:crypto";
+import { INGEST_RESEARCH_GRAPH_ID } from "../constants.js";
+import { isResearchSufficient } from "../shouldRefine.js";
+import { getGraphContext } from "./shared/getGraphContext.js";
 import { dispatchResearchBatch } from "./shared/dispatchSseCustom.js";
 import type { ResearchLoopStateType, ResearchLoopStateUpdate } from "../state.js";
-import type { ExitReason, ResearchBatch } from "../types.js";
+import type { ResearchBatch, ResearchLoopCompileExitReason } from "../types.js";
+
+/**
+ * Derive the loop exit reason from evaluation + graph ownership.
+ *
+ * @param state Current research-loop state after `evaluate_sufficiency`.
+ * @param graphId Owning graph id from {@link GraphContext}.
+ */
+export function resolveResearchExitReason(
+  state: ResearchLoopStateType,
+  graphId: string,
+): ResearchLoopCompileExitReason {
+  if (isResearchSufficient(state.lastEvaluation)) return "score_threshold";
+  return graphId === INGEST_RESEARCH_GRAPH_ID ? "max_iterations" : "safety_cap";
+}
 
 /**
  * `compile_batch` node — pure projection that freezes the current state into a
@@ -28,9 +45,9 @@ export async function compileBatch(
   state: ResearchLoopStateType,
   config: LangGraphRunnableConfig,
 ): Promise<ResearchLoopStateUpdate> {
+  const ctx = getGraphContext(config);
   const score = state.lastEvaluation?.score ?? null;
-  const exitReason: ExitReason =
-    score !== null && score >= 0.75 ? "score_threshold" : "max_iterations";
+  const exitReason = resolveResearchExitReason(state, ctx.graphId);
   const batch: ResearchBatch = {
     id: randomUUID(),
     iteration: state.iteration,
diff --git a/server/api/src/agents/subgraphs/research/nodes/evaluateSufficiency.ts b/server/api/src/agents/subgraphs/research/nodes/evaluateSufficiency.ts
index aefb2cae..181f76eb 100644
--- a/server/api/src/agents/subgraphs/research/nodes/evaluateSufficiency.ts
+++ b/server/api/src/agents/subgraphs/research/nodes/evaluateSufficiency.ts
@@ -6,7 +6,7 @@
  * 現在の `pendingSources` が brief を満たしているかを LLM で評価し、
  * `score` (0..1) と `missingAspects` を返す。post-increment した `iteration`
  * を返すことで、後段の `shouldRefine` がループ終了条件
- * (`score >= 0.75 || iteration >= maxIterations`) を正しく判定できる。
+ * (`score >= threshold || iteration >= cap`) を正しく判定できる。
  */
 import type { LangGraphRunnableConfig } from "@langchain/langgraph";
 import { z } from "zod";
@@ -15,19 +15,24 @@ import { createZediChatModel } from "../../../core/llm/modelFactory.js";
 import { resolveWikiComposeModelId } from "../../../core/llm/wikiComposeModelId.js";
 import { getGraphContext } from "./shared/getGraphContext.js";
 import { dispatchResearchEvaluation } from "./shared/dispatchSseCustom.js";
+import { RESEARCH_SUFFICIENCY_SCORE_THRESHOLD } from "../constants.js";
 import type { ResearchLoopStateType, ResearchLoopStateUpdate } from "../state.js";
 import type { Evaluation } from "../types.js";
 
 export const evaluationSchema = z.object({
   score: z.number().min(0).max(1),
+  sufficient: z.boolean(),
   rationale: z.string().min(1).max(500),
   missingAspects: z.array(z.string().min(1)).max(5),
 });
 
 const SYSTEM_PROMPT =
   "You are evaluating whether the research sources collected so far are sufficient " +
-  "to write the requested wiki article. Score 0..1 (≥0.75 means 'good enough'), " +
-  "give a short rationale, and list up to 5 missing aspects. Output JSON only.";
+  "to write the requested wiki article. Set `sufficient` to true when coverage is " +
+  "good enough to proceed without another research pass. Also provide a 0..1 score " +
+  `(>= ${RESEARCH_SUFFICIENCY_SCORE_THRESHOLD} typically means sufficient), a short ` +
+  "rationale, and up to 5 missing aspects that would still matter for the article. " +
+  "Output JSON only.";
 
 function buildUserPrompt(state: ResearchLoopStateType): string {
   const brief = state.messages
@@ -49,7 +54,8 @@ function buildUserPrompt(state: ResearchLoopStateType): string {
     `[Sources collected: ${state.pendingSources.length}]`,
     ...sourceLines,
     "",
-    `Iteration so far: ${state.iteration} / ${state.maxIterations}`,
+    `Research iteration completed so far: ${state.iteration}`,
+    "Continue refining only if important gaps remain for the brief.",
   ].join("\n");
 }
 
@@ -97,6 +103,7 @@ export async function evaluateSufficiency(
   ]);
   const evaluation: Evaluation = {
     score: parsed.score,
+    sufficient: parsed.sufficient,
     rationale: parsed.rationale,
     missingAspects: parsed.missingAspects,
   };
diff --git a/server/api/src/agents/subgraphs/research/nodes/planQueries.ts b/server/api/src/agents/subgraphs/research/nodes/planQueries.ts
index 1e730584..679e1a1d 100644
--- a/server/api/src/agents/subgraphs/research/nodes/planQueries.ts
+++ b/server/api/src/agents/subgraphs/research/nodes/planQueries.ts
@@ -2,7 +2,7 @@
  * `plan_queries` — generates the initial query set for the research loop.
  *
  * 調査ループの最初のノード。Brief / 指示メッセージから 1〜8 件の調査クエリを
- * 生成し、`maxIterations` を 1..5 にクランプする。"additional_research" 入力で
+ * 生成する。"additional_research" 入力で
  * 既存セッションの追加調査として呼ばれた場合、`iteration / lastEvaluation /
  * exitReason` をリセットし、`carryOverApprovedIds` で `pendingSources` を初期化
  * する（issue #949 の追加調査 API パス）。
@@ -52,12 +52,7 @@ const SYSTEM_PROMPT =
   "and 'web' for queries needing fresh public information. Output JSON only.";
 
 import type { AdditionalResearchRequest } from "../types.js";
-
-function clampMaxIterations(raw: unknown): number {
-  if (typeof raw !== "number" || !Number.isFinite(raw)) return 3;
-  const truncated = Math.trunc(raw);
-  return Math.min(Math.max(truncated, 1), 5);
-}
+import { resolveResearchMaxIterations } from "../constants.js";
 
 function briefFromState(
   state: ResearchLoopStateType,
@@ -97,9 +92,8 @@ export async function planQueries(
   const additional = state.additionalRequest ?? null;
   const brief = briefFromState(state, additional);
 
-  // Resolve maxIterations: input override > existing state > default(3); clamp 1..5.
-  // maxIterations は既存 state を優先しつつ 1..5 にクランプ。
-  const maxIterations = clampMaxIterations(state.maxIterations ?? 3);
+  // maxIterations: ingest uses caller cap; Wiki Compose always uses safety cap.
+  const maxIterations = resolveResearchMaxIterations(ctx.graphId, state.maxIterations);
 
   const modelId = await resolveWikiComposeModelId("orchestrator", ctx.tier, ctx.db);
   const model = await createZediChatModel({
diff --git a/server/api/src/agents/subgraphs/research/nodes/refineQueries.ts b/server/api/src/agents/subgraphs/research/nodes/refineQueries.ts
index 95738e13..f3a5ee7d 100644
--- a/server/api/src/agents/subgraphs/research/nodes/refineQueries.ts
+++ b/server/api/src/agents/subgraphs/research/nodes/refineQueries.ts
@@ -30,7 +30,7 @@ function buildUserPrompt(state: ResearchLoopStateType): string {
   const prior = state.queries.map((q) => `- ${q.query} (${q.channels.join("/")})`);
   const sourceTitles = state.pendingSources.map((s) => `- [${s.kind}] ${s.title}`);
   return [
-    `[Iteration ${state.iteration} / ${state.maxIterations}]`,
+    `[Research iteration ${state.iteration}]`,
     `Previous evaluation score: ${evaluation?.score ?? "n/a"}`,
     "",
     "[Missing aspects to address]",
diff --git a/server/api/src/agents/subgraphs/research/nodes/shared/dispatchSseCustom.ts b/server/api/src/agents/subgraphs/research/nodes/shared/dispatchSseCustom.ts
index 02da4b9a..92681a59 100644
--- a/server/api/src/agents/subgraphs/research/nodes/shared/dispatchSseCustom.ts
+++ b/server/api/src/agents/subgraphs/research/nodes/shared/dispatchSseCustom.ts
@@ -7,6 +7,7 @@
  */
 import { dispatchCustomEvent } from "@langchain/core/callbacks/dispatch";
 import type { LangGraphRunnableConfig } from "@langchain/langgraph";
+import type { ResearchLoopCompileExitReason } from "../../types.js";
 
 /** Payload shape for `research_iteration` custom events. */
 export interface ResearchIterationPayload {
@@ -29,7 +30,7 @@ export interface ResearchBatchPayload {
   iteration: number;
   sourceCount: number;
   score: number | null;
-  exitReason: "score_threshold" | "max_iterations";
+  exitReason: ResearchLoopCompileExitReason;
 }
 
 /**
diff --git a/server/api/src/agents/subgraphs/research/researchGraph.ts b/server/api/src/agents/subgraphs/research/researchGraph.ts
index bfb91d73..2dfe3f2d 100644
--- a/server/api/src/agents/subgraphs/research/researchGraph.ts
+++ b/server/api/src/agents/subgraphs/research/researchGraph.ts
@@ -5,7 +5,8 @@
  * `fetch_articles` → `evaluate_sufficiency` を 1 イテレーションとし、
  * `shouldRefine` の判定で `refine_queries` (= 次ループ) か `compile_batch` →
  * `human_review_research` (= HITL 中断) のいずれかに分岐する。終了条件:
- * `score >= 0.75` OR `iteration >= maxIterations` (default 3, clamp 1..5)。
+ * 評価 LLM が `score >= 0.75` と判断したとき、または安全上限
+ * {@link RESEARCH_SAFETY_MAX_ITERATIONS}（ingest が 1..5 を明示した場合はその cap）。
  *
  * Cyclic LangGraph with a parallel fan-out (`web_search ∥ wiki_search`) and a
  * conditional edge after `evaluate_sufficiency`. The HITL stop is implemented
@@ -78,9 +79,10 @@ export function registerResearchLoopGraph(): void {
     phase: "research",
     description:
       "Wiki Compose P1: autonomous research loop. Plans queries, runs web + wiki search, " +
-      "fetches articles, evaluates sufficiency, optionally refines and re-loops up to " +
-      "maxIterations (1..5, default 3), then interrupts at human_review_research for " +
-      "HITL source approval. Resume payload: { approvedSourceIds, rejectedSourceIds?, note? }.",
+      "fetches articles, evaluates sufficiency, optionally refines and re-loops until the " +
+      "evaluator LLM deems sources sufficient (score >= 0.75) or a safety cap is reached, " +
+      "then interrupts at human_review_research for HITL source approval. " +
+      "Resume payload: { approvedSourceIds, rejectedSourceIds?, note? }.",
     factory,
   });
 }
diff --git a/server/api/src/agents/subgraphs/research/shouldRefine.ts b/server/api/src/agents/subgraphs/research/shouldRefine.ts
index 1368b1ee..0523d09a 100644
--- a/server/api/src/agents/subgraphs/research/shouldRefine.ts
+++ b/server/api/src/agents/subgraphs/research/shouldRefine.ts
@@ -1,18 +1,34 @@
 /**
  * Research loop exit predicate (`evaluate_sufficiency` → refine | compile).
  */
+import { RESEARCH_SUFFICIENCY_SCORE_THRESHOLD } from "./constants.js";
 import type { ResearchLoopStateType } from "./state.js";
+import type { Evaluation } from "./types.js";
+
+export { RESEARCH_SUFFICIENCY_SCORE_THRESHOLD };
+
+/**
+ * Whether the latest evaluation considers research sufficient to compile.
+ *
+ * Prefers the evaluator's explicit `sufficient` flag; falls back to score threshold.
+ *
+ * @param evaluation Latest evaluation from `evaluate_sufficiency`, if any.
+ */
+export function isResearchSufficient(evaluation: Evaluation | null | undefined): boolean {
+  if (!evaluation) return false;
+  if (evaluation.sufficient) return true;
+  return evaluation.score >= RESEARCH_SUFFICIENCY_SCORE_THRESHOLD;
+}
 
 /**
  * 終了条件判定。`evaluate_sufficiency` の直後に呼ばれる。
  *
- * - `score >= 0.75` → `"compile"`
- * - `iteration >= maxIterations` → `"compile"`
+ * - evaluator marks sufficient → `"compile"`
+ * - `iteration >= maxIterations` → `"compile"` (ingest cap or safety cap)
  * - otherwise → `"refine"`
  */
 export function shouldRefine(state: ResearchLoopStateType): "refine" | "compile" {
-  const score = state.lastEvaluation?.score;
-  if (typeof score === "number" && score >= 0.75) return "compile";
+  if (isResearchSufficient(state.lastEvaluation)) return "compile";
   if (state.iteration >= state.maxIterations) return "compile";
   return "refine";
 }
diff --git a/server/api/src/agents/subgraphs/research/state.ts b/server/api/src/agents/subgraphs/research/state.ts
index 425459bf..174ef929 100644
--- a/server/api/src/agents/subgraphs/research/state.ts
+++ b/server/api/src/agents/subgraphs/research/state.ts
@@ -15,6 +15,7 @@
  */
 import { Annotation } from "@langchain/langgraph";
 import { BaseState } from "../../core/state/baseState.js";
+import { RESEARCH_SAFETY_MAX_ITERATIONS } from "./constants.js";
 import type {
   AdditionalResearchRequest,
   Evaluation,
@@ -61,10 +62,10 @@ export const ResearchLoopState = Annotation.Root({
     reducer: (_prev, next) => next,
     default: () => 0,
   }),
-  /** ループ回数上限（1..5、デフォルト 3）。`plan_queries` で clamp 確定。 */
+  /** ループ回数上限。ingest が 1..5 を明示した場合はその cap、それ以外は安全上限。 */
   maxIterations: Annotation<number>({
     reducer: (prev, next) => next ?? prev,
-    default: () => 3,
+    default: () => RESEARCH_SAFETY_MAX_ITERATIONS,
   }),
   /** 直近のクエリリスト。`plan_queries` / `refine_queries` が全置換する。 */
   queries: Annotation<PlannedQuery[]>({
diff --git a/server/api/src/agents/subgraphs/research/types.ts b/server/api/src/agents/subgraphs/research/types.ts
index 036df588..e07e8ae5 100644
--- a/server/api/src/agents/subgraphs/research/types.ts
+++ b/server/api/src/agents/subgraphs/research/types.ts
@@ -82,15 +82,18 @@ export interface PlannedQuery {
 }
 
 /**
- * `evaluate_sufficiency` ノードの出力。`score >= 0.75` で `compile_batch` 側へ
+ * `evaluate_sufficiency` ノードの出力。`sufficient === true` または
+ * `score >= {@link RESEARCH_SUFFICIENCY_SCORE_THRESHOLD}` で `compile_batch` 側へ
  * 分岐する（{@link ./researchGraph.ts} の `shouldRefine`）。
  *
  * Output of `evaluate_sufficiency`. The conditional edge uses
- * `score >= 0.75` as the exit predicate.
+ * {@link Evaluation.sufficient} first, then the score threshold.
  */
 export interface Evaluation {
-  /** 0..1. ≥ 0.75 → exit; otherwise refine. */
+  /** 0..1 confidence score; used when {@link sufficient} is absent. */
   score: number;
+  /** Explicit stop signal from the evaluator LLM. */
+  sufficient: boolean;
   /** Short natural-language rationale for the score. */
   rationale: string;
   /** Up to 5 short labels for what's still missing. */
@@ -120,13 +123,24 @@ export interface ResearchBatch {
 }
 
 /**
- * ループ終了理由。`compile_batch` で確定し、HITL に渡される。
+ * Reasons the research loop exited at `compile_batch` (emitted via SSE).
  *
- * Reason the loop exited; set by `compile_batch`.
+ * `compile_batch` で確定し SSE / HITL に渡される終了理由。
  */
-export type ExitReason =
+export type ResearchLoopCompileExitReason =
   | "score_threshold"
+  /** Ingest planner hit its explicit 1..5 iteration cap. */
   | "max_iterations"
+  /** Autonomous Wiki Compose hit {@link RESEARCH_SAFETY_MAX_ITERATIONS}. */
+  | "safety_cap";
+
+/**
+ * ループ終了理由。`compile_batch` で確定し、HITL に渡される。
+ *
+ * Reason the loop exited; set by `compile_batch` or orchestrator skip nodes.
+ */
+export type ExitReason =
+  | ResearchLoopCompileExitReason
   | "manual_stop"
   /** Orchestrator skipped the research loop after Brief (#953). */
   | "brief_skip";
diff --git a/server/api/src/routes/composeSessions.ts b/server/api/src/routes/composeSessions.ts
index 3763a8cf..3be84a3a 100644
--- a/server/api/src/routes/composeSessions.ts
+++ b/server/api/src/routes/composeSessions.ts
@@ -16,9 +16,9 @@
  *
  * `wiki-compose-research` (#949 / P1):
  * - `POST /run` body.input shapes:
- *   - Initial run: `{ messages?: [...], maxIterations?: number }` (or any
+ *   - Initial run: `{ messages?: [...] }` (or any
  *     object; the graph reads `state.messages` set by LangGraph from
- *     `body.input`).
+ *     `body.input`). Research depth is decided autonomously by the evaluator LLM.
  *   - Additional research (re-run on a *new* session of the same graph id):
  *     `{ kind: "additional_research", instruction: string, brief?: string,
  *        carryOverApprovedIds?: string[] }`
diff --git a/src/components/wikiCompose/ComposePanel.tsx b/src/components/wikiCompose/ComposePanel.tsx
index 481e1a3e..65e0fa28 100644
--- a/src/components/wikiCompose/ComposePanel.tsx
+++ b/src/components/wikiCompose/ComposePanel.tsx
@@ -51,11 +51,7 @@ export interface ComposePanelProps {
 
   activity: ComposeActivity[];
 
-  onSubmitBrief: (input: {
-    answers: BriefAnswer[];
-    appendToExisting?: boolean;
-    researchMaxIterations?: number;
-  }) => Promise<void>;
+  onSubmitBrief: (input: { answers: BriefAnswer[]; appendToExisting?: boolean }) => Promise<void>;
   onSubmitResearchApproval: (input: {
     approvedSourceIds: string[];
     rejectedSourceIds?: string[];
diff --git a/src/components/wikiCompose/DialogueSection.tsx b/src/components/wikiCompose/DialogueSection.tsx
index 5d9a0f20..cbe8c2f5 100644
--- a/src/components/wikiCompose/DialogueSection.tsx
+++ b/src/components/wikiCompose/DialogueSection.tsx
@@ -11,7 +11,7 @@
  */
 import React, { useMemo, useState } from "react";
 import { useTranslation } from "react-i18next";
-import { Button, Card, CardContent, CardHeader, CardTitle, Slider } from "@zedi/ui";
+import { Button, Card, CardContent, CardHeader, CardTitle } from "@zedi/ui";
 import { Sparkles, RefreshCw, ArrowRight } from "lucide-react";
 import type {
   BriefAnswer,
@@ -30,11 +30,7 @@ export interface DialogueSectionProps {
   outlineProposal: OutlineSection[];
   isStreaming: boolean;
   /** Brief submission. */
-  onSubmitBrief: (input: {
-    answers: BriefAnswer[];
-    appendToExisting?: boolean;
-    researchMaxIterations?: number;
-  }) => Promise<void>;
+  onSubmitBrief: (input: { answers: BriefAnswer[]; appendToExisting?: boolean }) => Promise<void>;
   /** Structure submission. */
   onSubmitOutline: (input: { sections: OutlineSection[] }) => Promise<void>;
 }
@@ -70,7 +66,6 @@ export const DialogueSection: React.FC<DialogueSectionProps> = ({
   const [appendToExisting, setAppendToExisting] = useState<boolean>(
     Boolean(pageSnapshot?.hasContent),
   );
-  const [maxIterations, setMaxIterations] = useState<number>(3);
   const [submitting, setSubmitting] = useState(false);
 
   const canSubmitBrief = useMemo(
@@ -139,31 +134,6 @@ export const DialogueSection: React.FC<DialogueSectionProps> = ({
           </Card>
         ) : null}
 
-        <Card>
-          <CardHeader className="pb-2">
-            <CardTitle className="text-muted-foreground text-xs tracking-wide uppercase">
-              {t("wikiCompose.brief.researchDepthTitle")}
-            </CardTitle>
-          </CardHeader>
-          <CardContent className="space-y-2">
-            <div className="text-muted-foreground flex items-center justify-between text-xs">
-              <span>{t("wikiCompose.brief.researchQuick")}</span>
-              <span data-testid="research-iterations-value" className="text-foreground font-medium">
-                {t("wikiCompose.brief.iterationCount", { count: maxIterations })}
-              </span>
-              <span>{t("wikiCompose.brief.researchDeep")}</span>
-            </div>
-            <Slider
-              data-testid="research-iterations-slider"
-              min={1}
-              max={5}
-              step={1}
-              value={[maxIterations]}
-              onValueChange={(v) => setMaxIterations(v[0] ?? 3)}
-            />
-          </CardContent>
-        </Card>
-
         <div className="flex justify-end">
           <Button
             data-testid="submit-brief"
@@ -174,7 +144,6 @@ export const DialogueSection: React.FC<DialogueSectionProps> = ({
                 await onSubmitBrief({
                   answers: Object.values(answers),
                   appendToExisting,
-                  researchMaxIterations: maxIterations,
                 });
               } finally {
                 setSubmitting(false);
diff --git a/src/hooks/wiki/useWikiComposeSession.ts b/src/hooks/wiki/useWikiComposeSession.ts
index a6e15a43..a46cc60b 100644
--- a/src/hooks/wiki/useWikiComposeSession.ts
+++ b/src/hooks/wiki/useWikiComposeSession.ts
@@ -64,11 +64,7 @@ export interface UseWikiComposeSessionArgs {
 /** Hook return shape. */
 export interface UseWikiComposeSessionReturn extends WikiComposeSessionState {
   start: () => Promise<void>;
-  submitBrief: (input: {
-    answers: BriefAnswer[];
-    appendToExisting?: boolean;
-    researchMaxIterations?: number;
-  }) => Promise<void>;
+  submitBrief: (input: { answers: BriefAnswer[]; appendToExisting?: boolean }) => Promise<void>;
   submitResearchApproval: (input: {
     approvedSourceIds: string[];
     rejectedSourceIds?: string[];
diff --git a/src/i18n/locales/en/wikiCompose.json b/src/i18n/locales/en/wikiCompose.json
index 65c86aeb..d687a4c6 100644
--- a/src/i18n/locales/en/wikiCompose.json
+++ b/src/i18n/locales/en/wikiCompose.json
@@ -36,11 +36,6 @@
     "existingContentTitle": "Page already has content",
     "append": "Append to existing content",
     "replace": "Replace existing content",
-    "researchDepthTitle": "Research depth",
-    "researchQuick": "1 (quick)",
-    "researchDeep": "5 (deep)",
-    "iterationCount_one": "{{count}} iteration",
-    "iterationCount_other": "{{count}} iterations",
     "startResearch": "Start research",
     "required": "required",
     "answerOptionsAria": "Answer options",
diff --git a/src/i18n/locales/ja/wikiCompose.json b/src/i18n/locales/ja/wikiCompose.json
index 88084e83..5dec4501 100644
--- a/src/i18n/locales/ja/wikiCompose.json
+++ b/src/i18n/locales/ja/wikiCompose.json
@@ -36,11 +36,6 @@
     "existingContentTitle": "ページに既存の本文があります",
     "append": "既存の本文に追記する",
     "replace": "既存の本文を置き換える",
-    "researchDepthTitle": "調査の深さ",
-    "researchQuick": "1（手早く）",
-    "researchDeep": "5（深く）",
-    "iterationCount_one": "{{count}} 回",
-    "iterationCount_other": "{{count}} 回",
     "startResearch": "調査を開始",
     "required": "必須",
     "answerOptionsAria": "回答の選択肢",
diff --git a/src/lib/wikiCompose/composeService.ts b/src/lib/wikiCompose/composeService.ts
index abfc6137..5b1a03f9 100644
--- a/src/lib/wikiCompose/composeService.ts
+++ b/src/lib/wikiCompose/composeService.ts
@@ -167,7 +167,7 @@ export async function runSession(input: {
 /**
  * Resume a compose session from an interrupt. The `resume` payload shape is
  * graph-specific:
- * - `human_review_brief` — `{ answers, appendToExisting?, researchMaxIterations? }`
+ * - `human_review_brief` — `{ answers, appendToExisting? }`
  * - `human_review_research` — `{ approvedSourceIds, rejectedSourceIds?, note? }`
  * - `human_review_outline` — `{ sections }`
  * - `conflict_resolution` — `{ acknowledged: true, note?: string }`
diff --git a/src/lib/wikiCompose/types.ts b/src/lib/wikiCompose/types.ts
index 0774685a..c3ba6020 100644
--- a/src/lib/wikiCompose/types.ts
+++ b/src/lib/wikiCompose/types.ts
@@ -234,7 +234,7 @@ export type ComposeSseEvent =
       iteration: number;
       sourceCount: number;
       score: number | null;
-      exitReason: "score_threshold" | "max_iterations";
+      exitReason: "score_threshold" | "max_iterations" | "safety_cap";
     }
   | {
       type: "compose_phase";