diff --git a/docs/design/708-smilegate-select-ai-oci-genai/README.md b/docs/design/708-smilegate-select-ai-oci-genai/README.md
index 75adf4a..24b0553 100644
--- a/docs/design/708-smilegate-select-ai-oci-genai/README.md
+++ b/docs/design/708-smilegate-select-ai-oci-genai/README.md
@@ -67,3 +67,46 @@ credential, region, compartment, object list와 metadata instructions를 복제
 형식을 기록한다. 품질·지연시간 결과를 검토하기 전에는
 `sg_game_extract_mentions`와 `sg_game_match_candidate`의 운영 프로파일을 바꾸지
 않는다.
+
+후보 모델 비교는 `sql/adb/114_sgmp_create_scope_chat_candidate_profiles.sql`과
+`sql/adb/115_sgmp_scope_chat_candidate_benchmark.sql`으로 수행한다. 후보는
+운영과 분리된 `SGMP_SCOPE_*` 프로파일로 생성한다. 각 모델은 `NONE`, `SINGLE`,
+`MULTI`, `ALL` 질문에서 다음을 기록한다.
+
+- 초 단위 Chat 응답시간
+- 설명문·Markdown 없이 raw JSON만 반환했는지
+- `scope_hint`와 `game_mentions` 수가 기대값과 일치하는지
+- 생성 또는 Chat 호출 오류
+
+## 게임명 추출과 벡터 점수 판정
+
+게임명과 조회 범위 추출은 짧은 JSON Chat 작업이므로 운영 프로파일
+`SGMP_POC_OCI_COHERE_COMMAND` (`cohere.command-latest`)을 사용한다.
+게임 카탈로그의 `ALIASES_JSON`에는 게임명, 영문명, 약칭, 게임 ID, prefix와
+운영 등록 별칭을 JSON 배열로 저장한다. 이 JSON 배열 전체를 게임당 하나의
+임베딩으로 생성한다. 추출된 각 게임명은 해당 게임 벡터의 cosine distance와
+`SG_GAME_SCOPE_POLICY`에 저장된 최대값을 비교해 `MATCHED` 또는 `UNMATCHED`로
+판정한다. 후보 동일성 확인을 위한 별도 LLM 호출은 사용하지 않는다.
+
+정책값은 코드가 아니라 DB 설정 테이블에 저장한다. 게임 카탈로그·임베딩 모델이
+변경되면 운영자가 그 기준값을 조정할 수 있다. `NONE`/`SINGLE`/`MULTI`/`ALL`
+계약, Few-shot NL2SQL 전달 형식, Text2SQL 프로파일은 유지한다.
+
+### 검증 기준
+
+- 추출 결과는 JSON 파싱 가능하고 `game_mentions`, `scope_hint`만 반환한다.
+- `NONE`, `SINGLE`, `MULTI`, `ALL`에서 canonical `targets`,
+  `dataEligibleTargets`, `unresolvedTargets`가 유지된다.
+- 동일성 판정을 위한 추가 OCI Chat 호출이 발생하지 않는다.
+
+### 후보 프로파일 비교 결과
+
+동일한 JSON 추출 프롬프트로 `NONE`, `SINGLE`, `MULTI`, `ALL`을 호출한 결과,
+`cohere.command-latest`, `cohere.command-plus-latest`, `cohere.command-a-vision`은
+기본 4건에서 모두 범위와 게임명 수를 맞췄다. `google.gemini-2.5-flash-lite`는
+JSON 파싱이 일관되지 않았고, Grok과 Llama Maverick은 게임 미지정 또는 전체 게임
+질의에서 범위 오류가 있었다.
+
+게임명 추출은 짧은 단일 JSON 요청이므로 `cohere.command-latest`를 선택했다.
+이 측정값은 ADB 내부 Chat 호출 시간이며, Portal의 MCP·ReAct·답변 합성 시간을
+포함한 전체 체감시간과는 별도로 관리한다.
diff --git a/docs/design/739-game-target-contract/README.md b/docs/design/739-game-target-contract/README.md
index db58402..b222ec4 100644
--- a/docs/design/739-game-target-contract/README.md
+++ b/docs/design/739-game-target-contract/README.md
@@ -105,9 +105,17 @@
 - 운영 registry에만 존재하는 게임도 카탈로그에 동기화하고 임베딩을 생성한다.
 - 게임별 물리 객체가 없으면 컬럼은 `null`이다.
 
+`SG_GAME_CATALOG.ALIASES_JSON`에는 게임명, 영문명, 약칭, 게임 ID, prefix와
+운영 registry 별칭을 JSON 배열로 저장한다. 배열 전체를 게임당 하나의 벡터로
+임베딩한다. `SG_GAME_SCOPE_POLICY.GAME_ALIAS_MAX_COSINE_DISTANCE`는 벡터 검색
+점수의 신뢰도 기준이며, 후보·점수·원 질문은 `queryPlan`에 함께 보존해 다음
+Select AI가 이름 근거를 보조 판단할 수 있게 한다. 별도 LLM 후보 동일성 판정은
+사용하지 않는다.
+
 `SG_GAME_CATALOG_SEARCH`는 물리 객체명을 후보 결과에 포함한다.
-`SG_GAME_QUERY_PLAN`은 추출된 mention마다 후보 선택 결과를 `targets`로 만들며,
-`ALL`은 벡터 선택 없이 전체 카탈로그를 사용한다.
+`SG_GAME_QUERY_PLAN`은 추출된 mention마다 벡터 점수 기반 후보 결과를 `targets`와
+`mentionResults.candidateGames`로 만들며, `ALL`은 벡터 선택 없이 전체 카탈로그를
+사용한다.
 
 ## 5. MCP와 NL2SQL 설계
 
diff --git a/sql/adb/114_sgmp_create_scope_chat_candidate_profiles.sql b/sql/adb/114_sgmp_create_scope_chat_candidate_profiles.sql
new file mode 100644
index 0000000..fce0828
--- /dev/null
+++ b/sql/adb/114_sgmp_create_scope_chat_candidate_profiles.sql
@@ -0,0 +1,90 @@
+-- Creates non-operational OCI profiles for game-scope extraction benchmarks.
+-- Every profile inherits the active GPT profile's OCI credential, region,
+-- object list, and metadata. Only model is varied.
+
+DECLARE
+  v_credential_name VARCHAR2(128);
+  v_region VARCHAR2(128);
+  v_compartment_id VARCHAR2(4000);
+  v_attributes CLOB;
+  v_attribute_json JSON_OBJECT_T;
+  v_exists PLS_INTEGER;
+
+  PROCEDURE create_candidate(
+    p_profile_name IN VARCHAR2,
+    p_model IN VARCHAR2
+  ) IS
+  BEGIN
+    SELECT COUNT(*)
+      INTO v_exists
+      FROM user_cloud_ai_profiles
+     WHERE profile_name = p_profile_name;
+
+    IF v_exists > 0 THEN
+      DBMS_CLOUD_AI.DROP_PROFILE(profile_name => p_profile_name, force => TRUE);
+    END IF;
+
+    v_attribute_json := JSON_OBJECT_T();
+    v_attribute_json.put('provider', 'oci');
+    v_attribute_json.put('credential_name', v_credential_name);
+    v_attribute_json.put('model', p_model);
+    v_attribute_json.put('region', v_region);
+    v_attribute_json.put('oci_compartment_id', v_compartment_id);
+    v_attributes := v_attribute_json.to_clob;
+
+    DBMS_CLOUD_AI.CREATE_PROFILE(
+      profile_name => p_profile_name,
+      attributes   => v_attributes,
+      description  => 'Non-operational Smilegate game-scope benchmark profile'
+    );
+
+    FOR source_attribute IN (
+      SELECT attribute_name, attribute_value
+        FROM user_cloud_ai_profile_attributes
+       WHERE profile_name = 'SGMP_POC_OCI_GPT54MINI'
+         AND attribute_name NOT IN (
+           'credential_name', 'model', 'provider', 'provider_endpoint',
+           'region', 'oci_compartment_id', 'oci_endpoint_id',
+           'oci_apiformat', 'oci_runtimetype'
+         )
+    ) LOOP
+      DBMS_CLOUD_AI.SET_ATTRIBUTE(
+        profile_name    => p_profile_name,
+        attribute_name  => source_attribute.attribute_name,
+        attribute_value => source_attribute.attribute_value
+      );
+    END LOOP;
+  END;
+BEGIN
+  SELECT DBMS_LOB.SUBSTR(attribute_value, 128, 1)
+    INTO v_credential_name
+    FROM user_cloud_ai_profile_attributes
+   WHERE profile_name = 'SGMP_POC_OCI_GPT54MINI'
+     AND attribute_name = 'credential_name';
+
+  SELECT DBMS_LOB.SUBSTR(attribute_value, 128, 1)
+    INTO v_region
+    FROM user_cloud_ai_profile_attributes
+   WHERE profile_name = 'SGMP_POC_OCI_GPT54MINI'
+     AND attribute_name = 'region';
+
+  SELECT DBMS_LOB.SUBSTR(attribute_value, 4000, 1)
+    INTO v_compartment_id
+    FROM user_cloud_ai_profile_attributes
+   WHERE profile_name = 'SGMP_POC_OCI_GPT54MINI'
+     AND attribute_name = 'oci_compartment_id';
+
+  create_candidate('SGMP_SCOPE_COHERE_VISION', 'cohere.command-a-vision');
+  create_candidate('SGMP_SCOPE_COHERE_COMMAND', 'cohere.command-latest');
+  create_candidate('SGMP_SCOPE_COHERE_PLUS', 'cohere.command-plus-latest');
+  create_candidate('SGMP_SCOPE_GEMINI_FLASH', 'google.gemini-2.5-flash-lite');
+  create_candidate('SGMP_SCOPE_LLAMA_MAV', 'meta.llama-4-maverick-17b-128e-instruct-fp8');
+  create_candidate('SGMP_SCOPE_GROK_NONR', 'xai.grok-4.20-non-reasoning');
+END;
+/
+
+SELECT profile_name, attribute_value AS model
+  FROM user_cloud_ai_profile_attributes
+ WHERE profile_name LIKE 'SGMP_SCOPE_%'
+   AND attribute_name = 'model'
+ ORDER BY profile_name;
diff --git a/sql/adb/115_sgmp_scope_chat_candidate_benchmark.sql b/sql/adb/115_sgmp_scope_chat_candidate_benchmark.sql
new file mode 100644
index 0000000..14e4614
--- /dev/null
+++ b/sql/adb/115_sgmp_scope_chat_candidate_benchmark.sql
@@ -0,0 +1,112 @@
+-- Read-only benchmark for candidate game-scope extraction profiles.
+-- Korean test questions use UTF-8 base64 reconstruction for SQLcl safety.
+
+set serveroutput on size unlimited
+prompt SG_SCOPE_PROFILE_BENCHMARK_LOADED
+
+DECLARE
+  TYPE t_case IS RECORD (
+    case_name VARCHAR2(12),
+    question CLOB,
+    expected_scope VARCHAR2(20),
+    expected_mentions PLS_INTEGER
+  );
+  TYPE t_cases IS TABLE OF t_case INDEX BY PLS_INTEGER;
+  v_cases t_cases;
+  v_prompt_prefix CLOB :=
+    'Extract only game-name mentions from the user question. '
+    || 'Metrics, acronyms, dates, filters, and database object or column names are not game names unless they are themselves an explicit game title. '
+    || 'When a title-like noun directly qualifies a game data request such as user master, character, sales, AU, NRU, server, or game log, preserve that noun as a game-name mention even when it is not in a catalog. '
+    || 'Do not discard an unknown title merely because it cannot be resolved. General scope words such as common, overall, all, total, or every are not game-name mentions unless they are part of an explicit title. '
+    || 'Return exactly one JSON object with keys game_mentions (array of strings) '
+    || 'and scope_hint (GLOBAL, SINGLE_GAME, MULTI_GAME, ALL_GAMES, UNKNOWN). '
+    || 'Do not resolve names to IDs and do not generate SQL. '
+    || 'Return raw JSON only: no prose, no Markdown, and no code fence. Question: ';
+  v_result CLOB;
+  v_json JSON_OBJECT_T;
+  v_started PLS_INTEGER;
+  v_elapsed NUMBER;
+  v_scope VARCHAR2(20);
+  v_mentions PLS_INTEGER;
+  v_raw_json VARCHAR2(5);
+
+  PROCEDURE run_case(
+    p_profile_name IN VARCHAR2,
+    p_case t_case
+  ) IS
+  BEGIN
+    v_started := DBMS_UTILITY.GET_TIME;
+    v_result := DBMS_CLOUD_AI.GENERATE(
+      prompt => v_prompt_prefix || p_case.question,
+      profile_name => p_profile_name,
+      action => 'chat'
+    );
+    v_elapsed := (DBMS_UTILITY.GET_TIME - v_started) / 100;
+    v_json := JSON_OBJECT_T.parse(v_result);
+    v_raw_json := 'TRUE';
+    v_scope := v_json.get_string('scope_hint');
+    v_mentions := v_json.get_array('game_mentions').get_size;
+    DBMS_OUTPUT.PUT_LINE(
+      p_profile_name || '|' || p_case.case_name
+      || '|seconds=' || TO_CHAR(v_elapsed, 'FM9990D00')
+      || '|raw_json=' || v_raw_json
+      || '|scope=' || NVL(v_scope, 'NULL')
+      || '|mentions=' || v_mentions
+      || '|expected=' || p_case.expected_scope || '/' || p_case.expected_mentions
+    );
+  EXCEPTION
+    WHEN OTHERS THEN
+      DBMS_OUTPUT.PUT_LINE(
+        p_profile_name || '|' || p_case.case_name
+        || '|seconds=' || TO_CHAR(v_elapsed, 'FM9990D00')
+        || '|ERROR=' || SQLCODE || '|' || SUBSTR(SQLERRM, 1, 180)
+      );
+  END;
+BEGIN
+  DBMS_OUTPUT.PUT_LINE('benchmark_started');
+  v_cases(1).case_name := 'NONE';
+  v_cases(1).question := utl_i18n.raw_to_char(utl_encode.base64_decode(
+    utl_raw.cast_to_raw('7KCE7LK0IOunpOy2nCDslYzroKTspJgu')), 'AL32UTF8');
+  v_cases(1).expected_scope := 'GLOBAL';
+  v_cases(1).expected_mentions := 0;
+
+  v_cases(2).case_name := 'SINGLE';
+  v_cases(2).question := utl_i18n.raw_to_char(utl_encode.base64_decode(
+    utl_raw.cast_to_raw('7Lm07KCc64KYIOy1nOyLoCBBVSDslYzroKTspJgu')), 'AL32UTF8');
+  v_cases(2).expected_scope := 'SINGLE_GAME';
+  v_cases(2).expected_mentions := 1;
+
+  v_cases(3).case_name := 'MULTI';
+  v_cases(3).question := utl_i18n.raw_to_char(utl_encode.base64_decode(
+    utl_raw.cast_to_raw('66Gc65Oc64KY7J247J20656RIOy5tOygnOuCmOydmCAyMDI264WEIDfsm5QgMTXsnbwgQVXrpbwg6rCB6rCBIOyVjOugpOykmC4=')), 'AL32UTF8');
+  v_cases(3).expected_scope := 'MULTI_GAME';
+  v_cases(3).expected_mentions := 2;
+
+  v_cases(4).case_name := 'ALL';
+  v_cases(4).question := utl_i18n.raw_to_char(utl_encode.base64_decode(
+    utl_raw.cast_to_raw('7KCE7LK0IOqyjOyehOydmCDrp6Tstpwg7JWM66Ck7KSYLg==')), 'AL32UTF8');
+  v_cases(4).expected_scope := 'ALL_GAMES';
+  v_cases(4).expected_mentions := 0;
+
+  FOR profile_row IN (
+    SELECT profile_name
+      FROM user_cloud_ai_profiles
+     WHERE profile_name IN (
+       'SGMP_POC_OCI_GPT54MINI',
+       'SGMP_SCOPE_COHERE_VISION',
+       'SGMP_SCOPE_COHERE_COMMAND',
+       'SGMP_SCOPE_COHERE_PLUS',
+       'SGMP_SCOPE_GEMINI_FLASH',
+       'SGMP_SCOPE_LLAMA_MAV',
+       'SGMP_SCOPE_GROK_NONR'
+     )
+     ORDER BY profile_name
+  ) LOOP
+    DBMS_OUTPUT.PUT_LINE('profile=' || profile_row.profile_name);
+    FOR i IN 1 .. 4 LOOP
+      run_case(profile_row.profile_name, v_cases(i));
+    END LOOP;
+  END LOOP;
+END;
+/
+prompt SG_SCOPE_PROFILE_BENCHMARK_COMPLETED
diff --git a/sql/adb/116_sg_game_catalog_alias_embeddings.sql b/sql/adb/116_sg_game_catalog_alias_embeddings.sql
new file mode 100644
index 0000000..5754854
--- /dev/null
+++ b/sql/adb/116_sg_game_catalog_alias_embeddings.sql
@@ -0,0 +1,122 @@
+-- Store all game-name variants as one JSON array per game and embed that JSON
+-- as the canonical game-search vector. No customer game name is hardcoded.
+
+DECLARE
+  v_column_count PLS_INTEGER;
+BEGIN
+  SELECT COUNT(*)
+    INTO v_column_count
+    FROM user_tab_columns
+   WHERE table_name = 'SG_GAME_CATALOG'
+     AND column_name = 'ALIASES_JSON';
+
+  IF v_column_count = 0 THEN
+    EXECUTE IMMEDIATE 'ALTER TABLE sg_game_catalog ADD (aliases_json CLOB)';
+  END IF;
+END;
+/
+
+UPDATE sg_game_catalog
+   SET aliases_json = '[]'
+ WHERE aliases_json IS NULL;
+/
+
+DECLARE
+  v_constraint_count PLS_INTEGER;
+BEGIN
+  SELECT COUNT(*)
+    INTO v_constraint_count
+    FROM user_constraints
+   WHERE table_name = 'SG_GAME_CATALOG'
+     AND constraint_name = 'SG_GAME_CATALOG_ALIASES_JS_CK';
+
+  IF v_constraint_count = 0 THEN
+    EXECUTE IMMEDIATE
+      'ALTER TABLE sg_game_catalog ADD CONSTRAINT sg_game_catalog_aliases_js_ck '
+      || 'CHECK (aliases_json IS JSON)';
+  END IF;
+END;
+/
+
+MERGE INTO sg_game_catalog c
+USING (
+  WITH source_alias AS (
+    SELECT game_id AS game_key, game_nm AS alias_name
+      FROM comn_game_alias_bas
+     WHERE use_yn = 'Y' AND game_nm IS NOT NULL
+    UNION ALL
+    SELECT game_id, game_alias_nm
+      FROM comn_game_alias_bas
+     WHERE use_yn = 'Y' AND game_alias_nm IS NOT NULL
+    UNION ALL
+    SELECT game_id, game_id
+      FROM comn_game_alias_bas
+     WHERE use_yn = 'Y' AND game_id IS NOT NULL
+    UNION ALL
+    SELECT game_id, game_prefix
+      FROM comn_game_alias_bas
+     WHERE use_yn = 'Y' AND game_prefix IS NOT NULL
+    UNION ALL
+    SELECT game_key, display_name
+      FROM sg_game_scope_registry
+     WHERE active_yn = 'Y' AND display_name IS NOT NULL
+    UNION ALL
+    SELECT game_key, game_alias
+      FROM sg_game_scope_registry
+     WHERE active_yn = 'Y' AND game_alias IS NOT NULL
+  ),
+  deduplicated_alias AS (
+    SELECT game_key, alias_name
+      FROM source_alias
+     WHERE TRIM(alias_name) IS NOT NULL
+     GROUP BY game_key, alias_name
+  )
+  SELECT game_key,
+         JSON_ARRAYAGG(alias_name ORDER BY alias_name RETURNING CLOB) AS aliases_json
+    FROM deduplicated_alias
+   GROUP BY game_key
+) s
+ON (c.game_key = s.game_key)
+WHEN MATCHED THEN UPDATE SET
+  c.aliases_json = s.aliases_json,
+  c.updated_at = SYSTIMESTAMP;
+/
+
+-- A game has one canonical vector made from its complete JSON alias array.
+UPDATE sg_game_catalog c
+   SET c.embedding = DBMS_VECTOR.UTL_TO_EMBEDDING(
+         c.aliases_json,
+         JSON(sg_qa_vector_params('search_document'))
+       ),
+       c.updated_at = SYSTIMESTAMP
+ WHERE c.active_yn = 'Y';
+/
+
+COMMENT ON COLUMN sg_game_catalog.aliases_json IS
+  'Canonical JSON string array of every game-name variant used as the embedding input.';
+COMMENT ON COLUMN sg_game_catalog.embedding IS
+  'One vector per game, generated from the complete aliases_json array.';
+/
+
+CREATE OR REPLACE FUNCTION sg_game_catalog_search(
+  p_question IN CLOB,
+  p_top_k IN PLS_INTEGER DEFAULT 5
+) RETURN SYS_REFCURSOR AUTHID DEFINER IS
+  v_query VECTOR;
+  v_result SYS_REFCURSOR;
+BEGIN
+  v_query := DBMS_VECTOR.UTL_TO_EMBEDDING(
+    p_question,
+    JSON(sg_qa_vector_params('search_query'))
+  );
+  OPEN v_result FOR
+    SELECT game_key, game_id, game_prefix, game_nm, game_alias_nm,
+           user_master_object_name,
+           VECTOR_DISTANCE(embedding, v_query, COSINE) AS cosine_distance
+      FROM sg_game_catalog
+     WHERE active_yn = 'Y' AND embedding IS NOT NULL
+     ORDER BY VECTOR_DISTANCE(embedding, v_query, COSINE), priority, game_key
+     FETCH FIRST LEAST(GREATEST(NVL(p_top_k, 5), 1), 20) ROWS ONLY;
+  RETURN v_result;
+END;
+/
diff --git a/sql/adb/117_sg_game_scope_policy.sql b/sql/adb/117_sg_game_scope_policy.sql
new file mode 100644
index 0000000..55e7a56
--- /dev/null
+++ b/sql/adb/117_sg_game_scope_policy.sql
@@ -0,0 +1,44 @@
+-- Customer-managed score policy for vector-only game identity resolution.
+
+BEGIN
+  EXECUTE IMMEDIATE q'[
+    CREATE TABLE sg_game_scope_policy (
+      policy_key   VARCHAR2(128) PRIMARY KEY,
+      number_value NUMBER,
+      text_value   VARCHAR2(4000),
+      description  VARCHAR2(1000) NOT NULL,
+      active_yn    CHAR(1) DEFAULT 'Y' NOT NULL,
+      updated_at   TIMESTAMP(6) DEFAULT SYSTIMESTAMP NOT NULL,
+      CONSTRAINT sg_game_scope_policy_active_ck CHECK (active_yn IN ('Y', 'N'))
+    )]';
+EXCEPTION
+  WHEN OTHERS THEN
+    IF SQLCODE != -955 THEN RAISE; END IF;
+END;
+/
+
+MERGE INTO sg_game_scope_policy t
+USING (
+  SELECT 'GAME_ALIAS_MAX_COSINE_DISTANCE' AS policy_key,
+         0.500000 AS number_value,
+         CAST(NULL AS VARCHAR2(4000)) AS text_value,
+         'Maximum cosine distance for accepting the closest independently embedded game alias.' AS description
+    FROM dual
+) s
+ON (t.policy_key = s.policy_key)
+WHEN MATCHED THEN UPDATE SET
+  t.description = s.description,
+  t.active_yn = 'Y',
+  t.updated_at = SYSTIMESTAMP
+WHEN NOT MATCHED THEN INSERT (
+  policy_key, number_value, text_value, description, active_yn
+) VALUES (
+  s.policy_key, s.number_value, s.text_value, s.description, 'Y'
+);
+/
+
+COMMENT ON TABLE sg_game_scope_policy IS
+  'Customer-managed game scope policy values; changing a value requires no application deployment.';
+COMMENT ON COLUMN sg_game_scope_policy.number_value IS
+  'Numeric policy value. GAME_ALIAS_MAX_COSINE_DISTANCE applies to the closest alias vector.';
+/
diff --git a/sql/adb/118_sgmp_scope_command_profile.sql b/sql/adb/118_sgmp_scope_command_profile.sql
new file mode 100644
index 0000000..68007bb
--- /dev/null
+++ b/sql/adb/118_sgmp_scope_command_profile.sql
@@ -0,0 +1,78 @@
+-- Operational OCI Cohere profile for short game-name and scope extraction.
+-- It inherits the active GPT profile's OCI credential, region, and metadata.
+
+DECLARE
+  v_credential_name VARCHAR2(128);
+  v_region VARCHAR2(128);
+  v_compartment_id VARCHAR2(4000);
+  v_attributes CLOB;
+  v_attribute_json JSON_OBJECT_T;
+  v_exists PLS_INTEGER;
+BEGIN
+  SELECT COUNT(*)
+    INTO v_exists
+    FROM user_cloud_ai_profiles
+   WHERE profile_name = 'SGMP_POC_OCI_COHERE_COMMAND';
+
+  IF v_exists > 0 THEN
+    DBMS_CLOUD_AI.DROP_PROFILE(
+      profile_name => 'SGMP_POC_OCI_COHERE_COMMAND',
+      force => TRUE
+    );
+  END IF;
+
+  SELECT DBMS_LOB.SUBSTR(attribute_value, 128, 1)
+    INTO v_credential_name
+    FROM user_cloud_ai_profile_attributes
+   WHERE profile_name = 'SGMP_POC_OCI_GPT54MINI'
+     AND attribute_name = 'credential_name';
+
+  SELECT DBMS_LOB.SUBSTR(attribute_value, 128, 1)
+    INTO v_region
+    FROM user_cloud_ai_profile_attributes
+   WHERE profile_name = 'SGMP_POC_OCI_GPT54MINI'
+     AND attribute_name = 'region';
+
+  SELECT DBMS_LOB.SUBSTR(attribute_value, 4000, 1)
+    INTO v_compartment_id
+    FROM user_cloud_ai_profile_attributes
+   WHERE profile_name = 'SGMP_POC_OCI_GPT54MINI'
+     AND attribute_name = 'oci_compartment_id';
+
+  v_attribute_json := JSON_OBJECT_T();
+  v_attribute_json.put('provider', 'oci');
+  v_attribute_json.put('credential_name', v_credential_name);
+  v_attribute_json.put('model', 'cohere.command-latest');
+  v_attribute_json.put('region', v_region);
+  v_attribute_json.put('oci_compartment_id', v_compartment_id);
+  v_attributes := v_attribute_json.to_clob;
+
+  DBMS_CLOUD_AI.CREATE_PROFILE(
+    profile_name => 'SGMP_POC_OCI_COHERE_COMMAND',
+    attributes => v_attributes,
+    description => 'Smilegate operational OCI Cohere Command profile for game scope extraction'
+  );
+
+  FOR source_attribute IN (
+    SELECT attribute_name, attribute_value
+      FROM user_cloud_ai_profile_attributes
+     WHERE profile_name = 'SGMP_POC_OCI_GPT54MINI'
+       AND attribute_name NOT IN (
+         'credential_name', 'model', 'provider', 'provider_endpoint',
+         'region', 'oci_compartment_id', 'oci_endpoint_id',
+         'oci_apiformat', 'oci_runtimetype'
+       )
+  ) LOOP
+    DBMS_CLOUD_AI.SET_ATTRIBUTE(
+      profile_name => 'SGMP_POC_OCI_COHERE_COMMAND',
+      attribute_name => source_attribute.attribute_name,
+      attribute_value => source_attribute.attribute_value
+    );
+  END LOOP;
+END;
+/
+
+SELECT profile_name, attribute_value AS model
+  FROM user_cloud_ai_profile_attributes
+ WHERE profile_name = 'SGMP_POC_OCI_COHERE_COMMAND'
+   AND attribute_name = 'model';
diff --git a/sql/adb/80_sg_game_mention_extract.sql b/sql/adb/80_sg_game_mention_extract.sql
index 871c7a7..795b887 100644
--- a/sql/adb/80_sg_game_mention_extract.sql
+++ b/sql/adb/80_sg_game_mention_extract.sql
@@ -1,20 +1,34 @@
--- OCI GenAI chat extraction. This function returns JSON only; it never chooses tables.
+-- One short OCI Chat call extracts game-name mentions. Identity resolution is
+-- performed later by catalog vector score, not by another LLM call.
 CREATE OR REPLACE FUNCTION sg_game_extract_mentions(p_question IN CLOB)
   RETURN CLOB AUTHID DEFINER
 IS
   v_prompt CLOB;
   v_result CLOB;
+  v_extract JSON_OBJECT_T;
 BEGIN
   v_prompt := 'Extract only game-name mentions from the user question. '
+    || 'Metrics, acronyms, dates, filters, and database object or column names are not game names unless they are themselves an explicit game title. '
+    || 'When a title-like noun directly qualifies a game data request such as user master, character, sales, AU, NRU, server, or game log, preserve that noun as a game-name mention even when it is not in a catalog. '
+    || 'Do not discard an unknown title merely because it cannot be resolved. General scope words such as common, overall, all, total, or every are not game-name mentions unless they are part of an explicit title. '
     || 'Return exactly one JSON object with keys game_mentions (array of strings) '
     || 'and scope_hint (GLOBAL, SINGLE_GAME, MULTI_GAME, ALL_GAMES, UNKNOWN). '
-    || 'Do not resolve names to IDs and do not generate SQL. Question: '
+    || 'Do not resolve names to IDs and do not generate SQL. '
+    || 'Return raw JSON only: no prose, no Markdown, and no code fence. Question: '
     || DBMS_LOB.SUBSTR(p_question, 4000, 1);
+
   v_result := DBMS_CLOUD_AI.GENERATE(
     prompt => v_prompt,
-    profile_name => 'SGMP_POC_OCI_GPT54MINI',
+    profile_name => 'SGMP_POC_OCI_COHERE_COMMAND',
     action => 'chat'
   );
+  v_extract := JSON_OBJECT_T.parse(v_result);
+  IF NOT v_extract.has('game_mentions') OR NOT v_extract.has('scope_hint') THEN
+    RAISE_APPLICATION_ERROR(
+      -20091,
+      'Game mention extraction must return game_mentions and scope_hint JSON keys.'
+    );
+  END IF;
   RETURN v_result;
 END;
 /
diff --git a/sql/adb/81_sg_game_query_plan.sql b/sql/adb/81_sg_game_query_plan.sql
index e3d42f0..412811f 100644
--- a/sql/adb/81_sg_game_query_plan.sql
+++ b/sql/adb/81_sg_game_query_plan.sql
@@ -1,32 +1,57 @@
--- OCI GenAI candidate selection. Vector distance only orders candidates; the
--- model may select one supplied game_key or explicitly reject all candidates.
+-- Vector-score candidate selection. The closest catalog vector is accepted
+-- only when it is within the customer-managed policy threshold.
 CREATE OR REPLACE FUNCTION sg_game_match_candidate(
   p_mention IN CLOB,
   p_candidates_json IN CLOB
 ) RETURN CLOB AUTHID DEFINER
 IS
-  v_prompt CLOB;
-  v_result CLOB;
+  v_candidates JSON_ARRAY_T;
+  v_candidate JSON_OBJECT_T;
+  v_result JSON_OBJECT_T := JSON_OBJECT_T();
+  v_threshold NUMBER;
+  v_distance NUMBER;
+  v_game_key VARCHAR2(128);
 BEGIN
-  v_prompt := 'Decide whether the extracted game-name mention refers to exactly '
-    || 'one game in the supplied candidate array. Candidate vector distance is '
-    || 'retrieval evidence only and must not establish identity. Compare the '
-    || 'mention with candidate gameName, aliases, gameId, and gamePrefix. '
-    || 'If one candidate clearly identifies the same game, return exactly '
-    || '{"status":"MATCHED","game_key":"candidate key","reason":"short reason"}. '
-    || 'If none clearly identifies the same game, return exactly '
-    || '{"status":"UNMATCHED","game_key":null,"reason":"short reason"}. '
-    || 'Never invent a game_key and never return a key absent from the supplied '
-    || 'candidate array. Mention: '
-    || DBMS_LOB.SUBSTR(p_mention, 1000, 1)
-    || ' Candidates: '
-    || DBMS_LOB.SUBSTR(p_candidates_json, 12000, 1);
-  v_result := DBMS_CLOUD_AI.GENERATE(
-    prompt => v_prompt,
-    profile_name => 'SGMP_POC_OCI_GPT54MINI',
-    action => 'chat'
-  );
-  RETURN v_result;
+  SELECT number_value
+    INTO v_threshold
+    FROM sg_game_scope_policy
+   WHERE policy_key = 'GAME_ALIAS_MAX_COSINE_DISTANCE'
+     AND active_yn = 'Y';
+
+  v_candidates := JSON_ARRAY_T.parse(p_candidates_json);
+  IF v_candidates.get_size = 0 THEN
+    v_result.put('status', 'UNMATCHED');
+    v_result.put_null('game_key');
+    v_result.put('reason', 'No active game catalog vector candidate was returned.');
+    RETURN v_result.to_clob;
+  END IF;
+
+  v_candidate := TREAT(v_candidates.get(0) AS JSON_OBJECT_T);
+  v_distance := v_candidate.get_number('cosineDistance');
+  v_game_key := v_candidate.get_string('gameKey');
+
+  IF v_distance <= v_threshold THEN
+    v_result.put('status', 'MATCHED');
+    v_result.put('game_key', v_game_key);
+    v_result.put(
+      'reason',
+      'Closest catalog vector distance '
+      || TO_CHAR(v_distance, 'FM0D000000')
+      || ' is within configured maximum '
+      || TO_CHAR(v_threshold, 'FM0D000000') || '.'
+    );
+  ELSE
+    v_result.put('status', 'UNMATCHED');
+    v_result.put_null('game_key');
+    v_result.put(
+      'reason',
+      'Closest catalog vector distance '
+      || TO_CHAR(v_distance, 'FM0D000000')
+      || ' exceeds configured maximum '
+      || TO_CHAR(v_threshold, 'FM0D000000') || '.'
+    );
+  END IF;
+  RETURN v_result.to_clob;
 END;
 /
 
@@ -255,14 +280,14 @@ BEGIN
 
       IF v_matched_candidate IS NOT NULL THEN
         v_mention_result.put('status', 'MATCHED');
-        v_mention_result.put('reasonCode', 'LLM_CANDIDATE_MATCH');
+        v_mention_result.put('reasonCode', 'VECTOR_SCORE_MATCH');
         v_mention_result.put('matchedGameKey', v_selected_key);
 
         v_target := JSON_OBJECT_T.parse(v_matched_candidate.to_clob);
         v_target.put('mention', v_mention);
         v_target.put('status', 'MATCHED');
         v_target.put('matchStatus', 'MATCHED');
-        v_target.put('reasonCode', 'LLM_CANDIDATE_MATCH');
+        v_target.put('reasonCode', 'VECTOR_SCORE_MATCH');
         v_target.put('reason', v_reason);
         v_matched_count := v_matched_count + 1;
 
@@ -282,14 +307,7 @@ BEGIN
         END IF;
       ELSE
         v_mention_result.put('status', 'UNMATCHED');
-        IF v_decision_status = 'MATCHED' THEN
-          v_mention_result.put(
-            'reasonCode',
-            'LLM_SELECTED_UNKNOWN_CANDIDATE'
-          );
-        ELSE
-          v_mention_result.put('reasonCode', 'LLM_NO_CANDIDATE_MATCH');
-        END IF;
+        v_mention_result.put('reasonCode', 'VECTOR_SCORE_OVER_THRESHOLD');
 
         v_target := JSON_OBJECT_T();
         v_target.put('mention', v_mention);
diff --git a/src/main/java/com/cloudhandson/vpdbackoffice/service/GameCatalogVectorService.java b/src/main/java/com/cloudhandson/vpdbackoffice/service/GameCatalogVectorService.java
index 4d57674..6b47a2b 100644
--- a/src/main/java/com/cloudhandson/vpdbackoffice/service/GameCatalogVectorService.java
+++ b/src/main/java/com/cloudhandson/vpdbackoffice/service/GameCatalogVectorService.java
@@ -18,8 +18,8 @@ import org.springframework.stereotype.Service;
 /**
  * MCP database gateway for the ADB-owned game query planning functions.
  *
- * <p>Game mention extraction, vector candidate retrieval, and OCI GenAI
- * candidate selection run inside {@code SG_GAME_QUERY_PLAN}. Java validates
+ * <p>Game mention extraction, JSON-alias vector retrieval, and score-based
+ * scope resolution run inside {@code SG_GAME_QUERY_PLAN}. Java validates
  * authentication and transports the structured JSON result without applying
  * game-name matching rules.</p>
  */