diff --git a/app/ai/client.py b/app/ai/client.py
index cd87f5b..60d1b82 100644
--- a/app/ai/client.py
+++ b/app/ai/client.py
@@ -260,23 +260,6 @@ class AIClient:
         cfg = self.ai.deep or self.ai.primary
         return await self._request(cfg, prompt, system=system)
 
-    async def call_classifier(self, prompt: str) -> str:
-        """answerability classifier (config ai.classifier, Mac mini 26B MLX).
-
-        private _request 직접 호출(classifier_service)을 봉인하는 public 진입점. gate 는
-        caller(classifier_service)가 acquire_mlx_gate 로 관리 — call_primary 와 동일한
-        caller-managed 계약(여기서 self-gate 하면 caller 와 double-acquire 데드락).
-        """
-        return await self._request(self.ai.classifier, prompt)
-
-    async def call_verifier(self, prompt: str) -> str:
-        """semantic verifier (config ai.verifier, Mac mini 26B MLX).
-
-        private _request 직접 호출(verifier_service)을 봉인. gate 는 caller(verifier_service)
-        가 관리(caller-managed — self-gate 금지).
-        """
-        return await self._request(self.ai.verifier, prompt)
-
     # ─── Legacy API (classify_worker 교체 시 제거 예정) ───────────────────
 
     async def classify(self, text: str, cfg=None) -> dict:
diff --git a/app/api/search.py b/app/api/search.py
index 5d6f559..c4fb1c7 100644
--- a/app/api/search.py
+++ b/app/api/search.py
@@ -3,42 +3,28 @@
 실제 검색 파이프라인(retrieval → fusion → rerank → diversity → confidence)
 은 `services/search/search_pipeline.py::run_search()` 로 분리되어 있다.
 이 파일은 다음만 담당:
-  - Pydantic 스키마 (SearchResult / SearchResponse / SearchDebug / DebugCandidate
-    / Citation / AskResponse / AskDebug)
+  - Pydantic 스키마 (SearchResult / SearchResponse / SearchDebug / DebugCandidate)
   - `/search` endpoint wrapper (run_search 호출 + logger + telemetry + 직렬화)
-  - `/ask` endpoint wrapper (Phase 3.3 에서 추가)
 """
 
-import asyncio
-import hmac
-import time
 from datetime import date
-from typing import Annotated, Literal
+from typing import Annotated
 
-from fastapi import APIRouter, BackgroundTasks, Depends, Header, Query
+from fastapi import APIRouter, BackgroundTasks, Depends, Query
 from fastapi.responses import JSONResponse
 from pydantic import BaseModel
 from sqlalchemy.ext.asyncio import AsyncSession
 
 from core.auth import get_current_user
-from core.config import settings
 from core.database import get_session
 from core.utils import setup_logger
 from models.user import User
-from services.document_telemetry import sanitize_source
-from services.search.classifier_service import ClassifierResult, classify
-from services.search.evidence_service import EvidenceItem, extract_evidence
 from services.search.fusion_service import DEFAULT_FUSION
-from services.search.grounding_check import check as grounding_check
-from services.search.refusal_gate import RefusalDecision, decide as refusal_decide
 from services.search import query_rewriter
 from services.search.retrieval_service import AxisFilter
 from services.search.result_decorate import compute_facets, decorate_version_status
 from services.search.search_pipeline import PipelineResult, run_search
-from services.search.synthesis_service import SynthesisResult, synthesize
-from services.search.verifier_service import VerifierResult, verify
-from services.prompt_versions import ASK_PROMPT_VERSION, resolve_primary_model
-from services.search_telemetry import record_ask_event, record_search_event
+from services.search_telemetry import record_search_event
 
 # logs/search.log + stdout 동시 출력 (Phase 0.4)
 logger = setup_logger("search")
@@ -354,241 +340,3 @@ async def search(
         debug=debug_obj,
         facets=facets_obj,
     )
-
-
-# ═══════════════════════════════════════════════════════════
-# Phase 3.3: /api/search/ask — Evidence + Grounded Synthesis
-# ═══════════════════════════════════════════════════════════
-
-
-class Citation(BaseModel):
-    """answer 본문의 [n] 에 해당하는 근거 단일 행."""
-
-    n: int
-    chunk_id: int | None
-    doc_id: int
-    title: str | None
-    section_title: str | None
-    span_text: str  # evidence LLM 이 추출한 50~300자
-    full_snippet: str  # 원본 800자 (citation 원문 보기 전용)
-    relevance: float
-    rerank_score: float
-
-
-class ConfirmedItem(BaseModel):
-    """Partial answer 의 개별 aspect 답변."""
-
-    aspect: str
-    text: str
-    citations: list[int]
-
-
-class AskDebug(BaseModel):
-    """`/ask?debug=true` 응답 확장."""
-
-    timing_ms: dict[str, float]
-    search_notes: list[str]
-    query_analysis: dict | None = None
-    confidence_signal: float
-    evidence_candidate_count: int
-    evidence_kept_count: int
-    evidence_skip_reason: str | None
-    synthesis_cache_hit: bool
-    synthesis_prompt_preview: str | None = None
-    synthesis_raw_preview: str | None = None
-    hallucination_flags: list[str] = []
-    # Phase 3.5a: per-layer defense 로깅
-    defense_layers: dict | None = None
-
-
-class AskResponse(BaseModel):
-    """`/ask` 응답. Phase 3.5a: completeness + aspects 추가."""
-
-    results: list[SearchResult]
-    ai_answer: str | None
-    citations: list[Citation]
-    synthesis_status: Literal[
-        "completed", "timeout", "skipped", "no_evidence", "parse_failed", "llm_error",
-        # PR-MacBook-RAG-Backend-1: 200 응답에는 등장하지 않음 (해당 status 는 503 분기).
-        # Literal 호환성 위해 포함.
-        "backend_unavailable",
-    ]
-    synthesis_ms: float
-    confidence: Literal["high", "medium", "low"] | None
-    refused: bool
-    no_results_reason: str | None
-    query: str
-    total: int
-    # Phase 3.5a
-    completeness: Literal["full", "partial", "insufficient"] = "full"
-    covered_aspects: list[str] | None = None
-    missing_aspects: list[str] | None = None
-    confirmed_items: list[ConfirmedItem] | None = None
-    # PR-MacBook-RAG-Backend-1: backend dispatcher metadata.
-    # backend 미지정 호출은 둘 다 None 으로 유지 (기존 호출자 호환 — Hermes docsrv_ask /
-    # voice-memo-bot 응답 형식 변동 0). 명시 opt-in 시만 채워짐.
-    backend_requested: str | None = None
-    backend_used: str | None = None
-    debug: AskDebug | None = None
-
-
-def _map_no_results_reason(
-    pr: PipelineResult,
-    evidence: list[EvidenceItem],
-    ev_skip: str | None,
-    sr: SynthesisResult,
-) -> str | None:
-    """사용자에게 보여줄 한국어 메시지 매핑.
-
-    Failure mode 표 (plan §Failure Modes) 기반.
-    """
-    # LLM 자가 refused → 모델이 준 사유 그대로
-    if sr.refused and sr.refuse_reason:
-        return sr.refuse_reason
-
-    # synthesis 상태 우선
-    if sr.status == "no_evidence":
-        if not pr.results:
-            return "검색 결과가 없습니다."
-        return "관련도 높은 근거를 찾지 못했습니다."
-    if sr.status == "skipped":
-        return "검색 결과가 없습니다."
-    if sr.status == "timeout":
-        return "답변 생성이 지연되어 생략했습니다. 검색 결과를 확인해 주세요."
-    if sr.status == "parse_failed":
-        return "답변 형식 오류로 생략했습니다."
-    if sr.status == "llm_error":
-        return "AI 서버에 일시적 문제가 있습니다."
-
-    # evidence 단계 실패는 fallback 을 탔더라도 notes 용
-    if ev_skip == "all_low_rerank":
-        return "관련도 높은 근거를 찾지 못했습니다."
-    if ev_skip == "empty_retrieval":
-        return "검색 결과가 없습니다."
-
-    return None
-
-
-def _build_citations(
-    evidence: list[EvidenceItem], used_citations: list[int]
-) -> list[Citation]:
-    """answer 본문에 실제로 등장한 n 만 Citation 으로 변환."""
-    by_n = {e.n: e for e in evidence}
-    out: list[Citation] = []
-    for n in used_citations:
-        e = by_n.get(n)
-        if e is None:
-            continue
-        out.append(
-            Citation(
-                n=e.n,
-                chunk_id=e.chunk_id,
-                doc_id=e.doc_id,
-                title=e.title,
-                section_title=e.section_title,
-                span_text=e.span_text,
-                full_snippet=e.full_snippet,
-                relevance=e.relevance,
-                rerank_score=e.rerank_score,
-            )
-        )
-    return out
-
-
-def _build_ask_debug(
-    pr: PipelineResult,
-    evidence: list[EvidenceItem],
-    ev_skip: str | None,
-    sr: SynthesisResult,
-    ev_ms: float,
-    synth_ms: float,
-    total_ms: float,
-) -> AskDebug:
-    timing: dict[str, float] = dict(pr.timing_ms)
-    timing["evidence_ms"] = ev_ms
-    timing["synthesis_ms"] = synth_ms
-    timing["ask_total_ms"] = total_ms
-
-    # candidate count 는 rule filter 통과한 수 (recomputable from results)
-    # 엄밀히는 evidence_service 내부 숫자인데, evidence 길이 ≈ kept, candidate
-    # 는 관측이 어려움 → kept 는 evidence 길이, candidate 는 별도 필드 없음.
-    # 단순화: candidate_count = len(evidence) 를 상한 근사로 둠 (debug 전용).
-    return AskDebug(
-        timing_ms=timing,
-        search_notes=pr.notes,
-        query_analysis=pr.query_analysis,
-        confidence_signal=pr.confidence_signal,
-        evidence_candidate_count=len(evidence),
-        evidence_kept_count=len(evidence),
-        evidence_skip_reason=ev_skip,
-        synthesis_cache_hit=sr.cache_hit,
-        synthesis_prompt_preview=None,  # 현재 synthesis_service 에서 노출 안 함
-        synthesis_raw_preview=sr.raw_preview,
-        hallucination_flags=sr.hallucination_flags,
-    )
-
-
-def _detect_synthesis_failure(sr: SynthesisResult) -> str | None:
-    """Synthesis 가 유효한 답을 못 냈으면 re_gate 라벨, 아니면 None.
-
-    판정 우선순위 (Phase 3.5 fix3):
-      1) sr.refused → LLM self-refuse (status="completed") 또는 mechanical fail 후 refused 전파
-         - status=="completed" + refused=True → "synthesis_self_refuse"
-         - 그 외 → f"synthesis_failed({status})"
-      2) sr.status ∈ {timeout, parse_failed, llm_error} → f"synthesis_failed({status})"
-      3) answer 공백 → f"synthesis_failed({status})"
-      4) 유효 → None
-    """
-    if sr.refused:
-        if sr.status == "completed":
-            return "synthesis_self_refuse"
-        return f"synthesis_failed({sr.status})"
-    if sr.status in ("timeout", "parse_failed", "llm_error"):
-        return f"synthesis_failed({sr.status})"
-    if not (sr.answer or "").strip():
-        return f"synthesis_failed({sr.status})"
-    return None
-
-
-def _resolve_eval_identity(
-    x_source: str | None,
-    x_eval_case_id: str | None,
-    x_eval_token: str | None,
-) -> tuple[str, str | None]:
-    """X-Source/X-Eval-Case-Id 신뢰 검증 (Phase 3.5 fix2).
-
-    규칙:
-      - 기본값: source='document_server', eval_case_id=None
-      - X-Source=eval 또는 X-Eval-Case-Id 가 들어왔다면 eval claim 으로 간주
-      - eval claim 은 X-Eval-Token == settings.eval_runner_token 일 때만 수용
-        (constant-time compare, env 미설정 시 항상 거부)
-      - 거부 시: 헤더 무시 + warning log + source=sanitize(non-eval) / eval_case_id=None
-      - 통과 시: source='eval', eval_case_id=x_eval_case_id
-
-    반환: (source, eval_case_id)
-    """
-    claimed_source = sanitize_source(x_source)
-    is_eval_claim = (claimed_source == "eval") or bool(x_eval_case_id)
-    if not is_eval_claim:
-        # 일반 호출 — eval_case_id 강제 None (source != 'eval' 이면 case_id 의미 없음)
-        return claimed_source, None
-
-    # eval claim — token 검증
-    expected = settings.eval_runner_token
-    presented = x_eval_token or ""
-    token_valid = bool(expected) and hmac.compare_digest(presented, expected)
-    if not token_valid:
-        logger.warning(
-            "eval header rejected: source=%s case_id=%s token_present=%s expected_set=%s",
-            x_source, x_eval_case_id, bool(x_eval_token), bool(expected),
-        )
-        # 일반 호출로 강등 — source='eval' 주장은 무시, case_id 도 무시
-        # claimed_source 가 'eval' 이면 default 'document_server' 로
-        if claimed_source == "eval":
-            return "document_server", None
-        return claimed_source, None
-
-    # token OK — eval 라벨 수용
-    return "eval", x_eval_case_id
-
-
diff --git a/app/prompts/classifier.txt b/app/prompts/classifier.txt
deleted file mode 100644
index 4f1a94e..0000000
--- a/app/prompts/classifier.txt
+++ /dev/null
@@ -1,33 +0,0 @@
-You are an answerability judge. Given a query and evidence chunks, determine if the evidence can answer the query. Respond ONLY in JSON.
-
-## CALIBRATION (CRITICAL)
-- verdict=full: evidence is SUFFICIENT to answer the CORE of the query. Missing minor details does NOT make it insufficient.
-- verdict=partial: evidence covers SOME major aspects but CLEARLY MISSES others the user explicitly asked about.
-- verdict=insufficient: evidence has NO relevant information for the query, or is completely off-topic.
-
-Example: Query="제6장 주요 내용", Evidence covers 제6장 definition+scope → verdict=full (core is covered).
-Example: Query="제6장 처벌 조항", Evidence covers 제6장 definition but NOT 처벌 → verdict=partial.
-Example: Query="감귤 출하량", Evidence about 산업안전보건법 → verdict=insufficient.
-
-## Rules
-1. Your "verdict" must be based ONLY on whether the CONTENT semantically answers the query. Ignore retrieval scores for this field.
-2. "covered_aspects": query aspects that evidence covers. Korean labels for Korean queries.
-3. "missing_aspects": query aspects that evidence does NOT cover. Korean labels.
-4. Keep aspects concise (2-5 words each), non-overlapping.
-
-## Output Schema
-{
-  "verdict": "full" | "partial" | "insufficient",
-  "covered_aspects": ["aspect1"],
-  "missing_aspects": ["aspect2"],
-  "confidence": "high" | "medium" | "low"
-}
-
-## Query
-{query}
-
-## Evidence chunks:
-{chunks}
-
-## Retrieval scores (for reference only, NOT for verdict):
-[{scores}]
diff --git a/app/prompts/verifier.txt b/app/prompts/verifier.txt
deleted file mode 100644
index aa3fa8b..0000000
--- a/app/prompts/verifier.txt
+++ /dev/null
@@ -1,42 +0,0 @@
-You are a grounding verifier. Given an answer and its evidence sources, check if the answer contradicts or fabricates information. Respond ONLY in JSON.
-
-## Contradiction Types (IMPORTANT — severity depends on type)
-- **direct_negation** (CRITICAL): Answer directly contradicts evidence. Examples: evidence "의무" but answer "권고"; evidence "금지" but answer "허용"; negation reversal ("~해야 한다" vs "~할 필요 없다").
-- **numeric_conflict**: Answer states a number different from evidence. "50명" in evidence but "100명" in answer. Only flag if the same concept is referenced. severity=critical when the number is the CORE answered quantity (amount/count/rate/date/duration that the query asked for); severity=minor when the number is peripheral (e.g., example/footnote).
-- **intent_core_mismatch**: Answer addresses a fundamentally different topic than the query asked about.
-- **nuance**: Answer overgeneralizes or adds qualifiers not in evidence (e.g., "모든" when evidence says "일부").
-- **unsupported_claim**: Answer makes a factual claim with no basis in any evidence.
-
-## Rules
-1. Compare each claim in the answer against the cited evidence. A claim with [n] citation should be checked against evidence [n].
-2. NOT a contradiction: Paraphrasing, summarizing, or restating the same fact in different words. Korean formal/informal style (합니다/한다) differences.
-3. Numbers must match exactly after normalization (1,000 = 1000). Range values (e.g., "100~200명") satisfy any answer within range.
-4. Legal/regulatory terms must preserve original meaning (의무 ≠ 권고, 금지 ≠ 제한, 허용 ≠ 금지).
-5. Maximum 5 contradictions (most severe first: direct_negation > numeric_conflict > intent_core_mismatch > nuance > unsupported_claim).
-
-## Output Schema
-{
-  "contradictions": [
-    {
-      "type": "direct_negation" | "numeric_conflict" | "intent_core_mismatch" | "nuance" | "unsupported_claim",
-      "severity": "critical" | "minor",
-      "claim": "answer 내 해당 구절 (50자 이내)",
-      "evidence_ref": "대응 근거 내용 (50자 이내, [n] 포함)",
-      "explanation": "모순 이유 (한국어, 30자 이내)"
-    }
-  ],
-  "verdict": "clean" | "minor_issues" | "major_issues"
-}
-
-severity mapping:
-- direct_negation → "critical"
-- numeric_conflict → "critical" if the number is the CORE answered quantity, else "minor"
-- All other types → "minor"
-
-If no contradictions: {"contradictions": [], "verdict": "clean"}
-
-## Answer
-{answer}
-
-## Evidence
-{numbered_evidence}
diff --git a/app/services/search/classifier_service.py b/app/services/search/classifier_service.py
deleted file mode 100644
index da44f48..0000000
--- a/app/services/search/classifier_service.py
+++ /dev/null
@@ -1,156 +0,0 @@
-"""Answerability classifier (Phase 3.5a).
-
-Mac mini 26B MLX 기반 (config.yaml ai.models.classifier — PR #20 이후 triage/primary/classifier 동일 endpoint). MLX gate 밖 — evidence extraction 과 병렬 실행 (concurrent 안전성 별 검토).
-
-P1 실측 결과: ternary (full/partial/insufficient) 불안정 → **binary (sufficient/insufficient)**.
-"full" vs "partial" 구분은 grounding_check 의 intent alignment 이 담당.
-
-Classifier verdict 는 "relevant evidence 가 있나" 의 binary 판단.
-covered_aspects / missing_aspects 는 로깅용으로 유지 (refusal gate 에서 사용 안 함).
-"""
-
-from __future__ import annotations
-
-import asyncio
-import time
-from dataclasses import dataclass
-from typing import Literal
-
-from ai.client import AIClient, _load_prompt, parse_json_response
-from core.config import settings
-from core.utils import setup_logger
-
-from .llm_gate import Priority, acquire_mlx_gate
-
-logger = setup_logger("classifier")
-
-LLM_TIMEOUT_MS = 30000
-CIRCUIT_THRESHOLD = 5
-CIRCUIT_RECOVERY_SEC = 60
-
-_failure_count = 0
-_circuit_open_until: float | None = None
-
-
-@dataclass(slots=True)
-class ClassifierResult:
-    status: Literal["ok", "timeout", "error", "circuit_open", "skipped"]
-    verdict: Literal["sufficient", "insufficient"] | None
-    covered_aspects: list[str]
-    missing_aspects: list[str]
-    elapsed_ms: float
-
-
-try:
-    CLASSIFIER_PROMPT = _load_prompt("classifier.txt")
-except FileNotFoundError:
-    CLASSIFIER_PROMPT = ""
-    logger.warning("classifier.txt not found — classifier will always skip")
-
-
-def _build_input(
-    query: str,
-    top_chunks: list[dict],
-    rerank_scores: list[float],
-) -> str:
-    """Y+ input (content + scores with role separation)."""
-    chunk_block = "\n".join(
-        f"[{i+1}] title: {c.get('title','')}\n"
-        f"    section: {c.get('section','')}\n"
-        f"    snippet: {c.get('snippet','')}"
-        for i, c in enumerate(top_chunks[:3])
-    )
-    scores_str = ", ".join(f"{s:.2f}" for s in rerank_scores[:3])
-    return (
-        CLASSIFIER_PROMPT
-        .replace("{query}", query)
-        .replace("{chunks}", chunk_block)
-        .replace("{scores}", scores_str)
-    )
-
-
-async def classify(
-    query: str,
-    top_chunks: list[dict],
-    rerank_scores: list[float],
-) -> ClassifierResult:
-    """Always-on binary classifier. Parallel with evidence extraction.
-
-    Returns:
-        ClassifierResult with verdict=sufficient|insufficient.
-        Status "ok" 이 아니면 verdict=None (caller 가 fallback 처리).
-    """
-    global _failure_count, _circuit_open_until
-    t_start = time.perf_counter()
-
-    # Circuit breaker
-    if _circuit_open_until and time.time() < _circuit_open_until:
-        return ClassifierResult("circuit_open", None, [], [], 0.0)
-
-    if not CLASSIFIER_PROMPT:
-        return ClassifierResult("skipped", None, [], [], 0.0)
-
-    if not hasattr(settings.ai, "classifier") or settings.ai.classifier is None:
-        return ClassifierResult("skipped", None, [], [], 0.0)
-
-    prompt = _build_input(query, top_chunks, rerank_scores)
-    client = AIClient()
-    try:
-        # 2026-05-17: PR #20 이후 endpoint 가 Mac mini 26B → llm_gate Semaphore(1) 필수.
-        # Gate 미사용 시 classifier + evidence + synthesis 가 동시에 single-inference
-        # MLX 에 race → 거의 모두 timeout (실측: 8/10 fixture query). docstring 영구 룰:
-        # "MLX primary 호출 경로는 예외 없이 gate 획득 필수".
-        async with acquire_mlx_gate(Priority.FOREGROUND):
-            async with asyncio.timeout(LLM_TIMEOUT_MS / 1000):
-                raw = await client.call_classifier(prompt)
-        _failure_count = 0
-    except asyncio.TimeoutError:
-        _failure_count += 1
-        if _failure_count >= CIRCUIT_THRESHOLD:
-            _circuit_open_until = time.time() + CIRCUIT_RECOVERY_SEC
-            logger.error(f"classifier circuit OPEN for {CIRCUIT_RECOVERY_SEC}s")
-        logger.warning("classifier timeout")
-        return ClassifierResult(
-            "timeout", None, [], [],
-            (time.perf_counter() - t_start) * 1000,
-        )
-    except Exception as e:
-        _failure_count += 1
-        if _failure_count >= CIRCUIT_THRESHOLD:
-            _circuit_open_until = time.time() + CIRCUIT_RECOVERY_SEC
-            logger.error(f"classifier circuit OPEN for {CIRCUIT_RECOVERY_SEC}s")
-        logger.warning("classifier error: type=%s repr=%r", type(e).__name__, e)
-        return ClassifierResult(
-            "error", None, [], [],
-            (time.perf_counter() - t_start) * 1000,
-        )
-    finally:
-        await client.close()
-
-    elapsed_ms = (time.perf_counter() - t_start) * 1000
-    parsed = parse_json_response(raw)
-    if not isinstance(parsed, dict):
-        logger.warning("classifier parse failed raw=%r", (raw or "")[:200])
-        return ClassifierResult("error", None, [], [], elapsed_ms)
-
-    # ternary → binary 매핑
-    raw_verdict = parsed.get("verdict", "")
-    if raw_verdict == "insufficient":
-        verdict: Literal["sufficient", "insufficient"] | None = "insufficient"
-    elif raw_verdict in ("full", "partial", "sufficient"):
-        verdict = "sufficient"
-    else:
-        verdict = None
-
-    covered = parsed.get("covered_aspects") or []
-    missing = parsed.get("missing_aspects") or []
-    if not isinstance(covered, list):
-        covered = []
-    if not isinstance(missing, list):
-        missing = []
-
-    logger.info(
-        "classifier ok query=%r verdict=%s (raw=%s) covered=%d missing=%d elapsed_ms=%.0f",
-        query[:60], verdict, raw_verdict, len(covered), len(missing), elapsed_ms,
-    )
-    return ClassifierResult("ok", verdict, covered, missing, elapsed_ms)
diff --git a/app/services/search/grounding_check.py b/app/services/search/grounding_check.py
deleted file mode 100644
index 66108ca..0000000
--- a/app/services/search/grounding_check.py
+++ /dev/null
@@ -1,505 +0,0 @@
-"""Grounding check — post-synthesis 검증 (Phase 3.5a).
-
-Strong/weak flag 분리:
-- **Strong** (→ partial 강등 or refuse): fabricated_number, intent_misalignment(important)
-- **Weak** (→ confidence lower only): uncited_claim, low_overlap, intent_misalignment(generic)
-
-Re-gate 로직 (Phase 3.5a 9라운드 토론 결과):
-- strong 1개 → partial 강등
-- strong 2개 이상 → refuse
-- weak → confidence "low" 만
-
-Intent alignment (rule-based):
-- query 의 핵심 명사가 answer 에 등장하는지 확인
-- "처벌" 같은 중요 키워드 누락은 strong
-- "주요", "관련" 같은 generic 은 무시
-"""
-
-from __future__ import annotations
-
-import re
-from dataclasses import dataclass
-from typing import TYPE_CHECKING
-
-from core.utils import setup_logger
-
-if TYPE_CHECKING:
-    from .evidence_service import EvidenceItem
-
-logger = setup_logger("grounding")
-
-# "주요", "관련" 등 intent alignment 에서 제외할 generic 단어
-GENERIC_TERMS = frozenset({
-    "주요", "관련", "내용", "정의", "기준", "방법", "설명", "개요",
-    "대한", "위한", "대해", "무엇", "어떤", "어떻게", "있는",
-    "하는", "되는", "이런", "그런", "이것", "그것",
-})
-
-
-@dataclass(slots=True)
-class GroundingResult:
-    strong_flags: list[str]
-    weak_flags: list[str]
-
-
-_UNIT_CHARS = r'명인개%년월일조항호세건원회'
-
-# "이상/이하/초과/미만" — threshold 표현 (numeric conflict 에서 skip 대상)
-_THRESHOLD_SUFFIXES = re.compile(r'이상|이하|초과|미만')
-
-# 약칭/근사치 prefix — 매칭 전 제거 (Phase 3.5 B1).
-# ⚠ 최대/최소 는 의도적으로 제외 — 이들은 bound operator 라 의미가 다름 (Phase 3.5 B1 fix3).
-# 약/대략/거의/얼추 만 노이즈 prefix 로 strip.
-_APPROX_PREFIX_RE = re.compile(r'(약|대략|거의|얼추)\s*')
-
-# 단위 동의어 dict — 추출 직후 정규화 (Phase 3.5 B1)
-# 의미가 동일한 단위는 같은 표기로 통일해서 set 비교/range overlap 안정화.
-_UNIT_SYNONYMS: dict[str, str] = {
-    "인": "명",
-    "사람": "명",
-    "퍼센트": "%",
-    "프로": "%",
-    "KRW": "원",
-    "krw": "원",
-}
-
-# tolerance(±1%) 허용 단위 — 양적 측정값 (Phase 3.5 B1)
-_TOLERANCE_UNITS: frozenset[str] = frozenset({"명", "원", "%", "건", "개"})
-
-# tolerance 미적용 단위 — 식별자성 숫자 (연도/조문/횟수)
-_EXACT_ONLY_UNITS: frozenset[str] = frozenset({"년", "월", "일", "조", "항", "호", "회"})
-
-# 최대/최소 prefix 패턴 — bound operator (Phase 3.5 B1 fix3).
-# 매칭된 숫자는 exact pool 에서 제외하고 one-sided range 로 변환.
-# 경계값 자체는 clear 대상 아님 (Codex 권장: "최대 100명" + answer "100명" → flag 유지).
-_BOUND_PATTERN_RE = re.compile(
-    rf'(최대|최소)\s*(\d[\d,.]*)\s*([{_UNIT_CHARS}]|인|사람|퍼센트|프로|KRW|krw)'
-)
-_RANGE_INF = 10**18  # one-sided range 상한 sentinel
-
-
-def _normalize_unit(unit: str) -> str:
-    """단위 동의어 → 대표 표기."""
-    return _UNIT_SYNONYMS.get(unit, unit)
-
-
-def _extract_unit(literal: str) -> str | None:
-    """리터럴에서 숫자 뒤 단위(한 글자 또는 동의어) 추출 + 정규화."""
-    # 천단위 콤마 + 옵션 소수 + 한글 단위 한 글자 또는 동의어
-    m = re.match(rf'[\d,.]+\s*([{_UNIT_CHARS}]|인|사람|퍼센트|프로|KRW|krw)', literal)
-    if not m:
-        return None
-    return _normalize_unit(m.group(1))
-
-
-def _extract_numeric_corpus(text: str) -> dict:
-    """단위별 숫자 + 범위 + bound 통합 추출 (Phase 3.5 B1 fix1+fix3).
-
-    Returns:
-      {
-        "exact_by_unit": {unit_or_None: set(digits)},   # 평범한 숫자 (bound 제외)
-        "ranges_by_unit": {unit: [(lo, hi), ...]},       # 양방향(A~B) + 단방향(최대/최소)
-      }
-
-    None 키는 단위 없는 bare 숫자.
-    `최대 N <unit>` → ranges[(0, N-1)] (경계값 자체는 cleared 대상 아님)
-    `최소 N <unit>` → ranges[(N+1, INF)]
-    """
-    cleaned = _APPROX_PREFIX_RE.sub('', text)
-
-    exact_by_unit: dict[str | None, set[str]] = {None: set()}
-    ranges_by_unit: dict[str, list[tuple[int, int]]] = {}
-
-    # 1) 최대/최소 — bound. exact pool 에서 제외, one-sided range 로 변환.
-    bound_spans: list[tuple[int, int]] = []  # 매칭 substring 위치 — 이후 단계에서 skip
-    for m in _BOUND_PATTERN_RE.finditer(cleaned):
-        bound_kind = m.group(1)
-        try:
-            n = int(m.group(2).replace(',', '').split('.')[0])
-        except ValueError:
-            continue
-        unit = _normalize_unit(m.group(3))
-        if bound_kind == "최대":
-            ranges_by_unit.setdefault(unit, []).append((0, max(0, n - 1)))
-        else:  # 최소
-            ranges_by_unit.setdefault(unit, []).append((n + 1, _RANGE_INF))
-        bound_spans.append((m.start(), m.end()))
-
-    def _in_bound_span(pos: int) -> bool:
-        return any(s <= pos < e for s, e in bound_spans)
-
-    # 2) 천단위 콤마 bare number
-    for m in re.finditer(r'\d{1,3}(?:,\d{3})+(?:\.\d+)?', cleaned):
-        if _in_bound_span(m.start()):
-            continue
-        exact_by_unit[None].add(m.group().replace(',', ''))
-
-    # 3) 단위 있는 숫자 (단위 동의어 포함)
-    for m in re.finditer(
-        rf'(\d[\d,.]*)\s*([{_UNIT_CHARS}]|인|사람|퍼센트|프로|KRW|krw)',
-        cleaned,
-    ):
-        if _in_bound_span(m.start()):
-            continue
-        digits = m.group(1).replace(',', '').split('.')[0]
-        if not digits:
-            continue
-        unit = _normalize_unit(m.group(2))
-        exact_by_unit.setdefault(unit, set()).add(digits)
-
-    # 4) 양방향 범위 표현 (A~B / A 부터 B)
-    for m in re.finditer(
-        rf'(\d[\d,.]*)\s*(?:[~\-–]|부터)\s*(\d[\d,.]*)\s*([{_UNIT_CHARS}]|인|사람|퍼센트|프로)',
-        cleaned,
-    ):
-        if _in_bound_span(m.start()):
-            continue
-        try:
-            lo = int(m.group(1).replace(',', '').split('.')[0])
-            hi = int(m.group(2).replace(',', '').split('.')[0])
-        except ValueError:
-            continue
-        unit = _normalize_unit(m.group(3))
-        ranges_by_unit.setdefault(unit, []).append((min(lo, hi), max(lo, hi)))
-
-    # 5) bare 2자리+ 단독 숫자
-    for m in re.finditer(r'\b(\d{2,})\b', cleaned):
-        if _in_bound_span(m.start()):
-            continue
-        exact_by_unit[None].add(m.group())
-
-    return {
-        "exact_by_unit": exact_by_unit,
-        "ranges_by_unit": ranges_by_unit,
-    }
-
-
-def _within_unit_range(
-    n: int, unit: str | None, ranges_by_unit: dict[str, list[tuple[int, int]]]
-) -> bool:
-    """unit-matching range 검증.
-
-    answer unit 이 None (bare 숫자) 면 보수적으로 False — bare 답변은 range clear 대상 아님.
-    """
-    if unit is None:
-        return False
-    return any(lo <= n <= hi for lo, hi in ranges_by_unit.get(unit, []))
-
-
-def _close_to_unit_pool(
-    n: int, unit: str | None, exact_by_unit: dict[str | None, set[str]], tol: float
-) -> bool:
-    """unit-matching tolerance 검증.
-
-    answer unit 이 None 이면 False — bare 답변은 tolerance 대상 아님.
-    같은 unit bucket 안의 후보만 비교.
-    """
-    if unit is None:
-        return False
-    candidates = exact_by_unit.get(unit, set())
-    for c in candidates:
-        try:
-            cn = int(c)
-        except ValueError:
-            continue
-        if cn == 0:
-            continue
-        if abs(n - cn) / cn <= tol:
-            return True
-    return False
-
-
-def _extract_number_literals(text: str) -> set[str]:
-    """숫자 + 단위 추출 + normalize (Phase 3.5 B1: 6단계 확장).
-
-    1) 약칭 prefix 제거 ("약 100명" → "100명")
-    2) 천단위 콤마 bare number 우선 ("1,000" → "1000" set 등록)
-    3) 한국어 단위 접미사 매칭 (기존)
-    4) 범위 표현 양쪽 숫자 추출 (separator: ~, -, –, 부터)
-    5) 단위 동의어 정규화 (인→명, 퍼센트→%, KRW→원)
-    6) bare 2자리+ 추출 (기존)
-    """
-    # 1. 약칭 prefix 제거 (전체 텍스트에서)
-    cleaned = _APPROX_PREFIX_RE.sub('', text)
-
-    # 2. 천단위 콤마 bare number — normalize 된 값을 set 에 선등록
-    normalized: set[str] = set()
-    for m in re.finditer(r'\d{1,3}(?:,\d{3})+(?:\.\d+)?', cleaned):
-        normalized.add(m.group().replace(',', ''))
-
-    # 3. 숫자 + 한국어 단위 접미사 (동의어 포함)
-    raw: set[str] = set(re.findall(
-        rf'\d[\d,.]*\s*(?:[{_UNIT_CHARS}]|인|사람|퍼센트|프로|KRW|krw)\w{{0,2}}',
-        cleaned,
-    ))
-
-    # 4. 범위 표현 — separator 에 "부터" 추가
-    for m in re.finditer(
-        rf'(\d[\d,.]*)\s*(?:[~\-–]|부터)\s*(\d[\d,.]*)\s*([{_UNIT_CHARS}]|인|사람|퍼센트|프로)',
-        cleaned,
-    ):
-        unit_norm = _normalize_unit(m.group(3))
-        raw.add(m.group(1) + unit_norm)
-        raw.add(m.group(2) + unit_norm)
-
-    # 5. normalize: 단위 동의어 통일 + 콤마 제거
-    for r in raw:
-        # 단위 부분 정규화
-        m = re.match(r'([\d,.]+)\s*([^\d\s]+)', r)
-        if m:
-            digits_part = m.group(1)
-            unit_part = _normalize_unit(m.group(2))
-            normalized.add(digits_part + unit_part)
-            normalized.add(digits_part.replace(',', '') + unit_part)
-        normalized.add(r.strip())
-        num_only = re.match(r'[\d,.]+', r)
-        if num_only:
-            normalized.add(num_only.group().replace(',', ''))
-
-    # 6. 단독 숫자 (2자리+ 만)
-    for d in re.findall(r'\b(\d{2,})\b', cleaned):
-        normalized.add(d)
-    return normalized
-
-
-def _within_evidence_range(digits: str, raw: str, evidence_text: str) -> bool:
-    """evidence 에 'A~B 단위' 가 있고 answer 의 숫자가 그 범위 안이면 True.
-
-    범위 단위는 무시 (단위 비교는 호출 전 단계). digits = 정수 문자열.
-    """
-    try:
-        n = int(digits)
-    except ValueError:
-        return False
-    cleaned_ev = _APPROX_PREFIX_RE.sub('', evidence_text)
-    for m in re.finditer(
-        rf'(\d[\d,.]*)\s*(?:[~\-–]|부터)\s*(\d[\d,.]*)\s*[{_UNIT_CHARS}]',
-        cleaned_ev,
-    ):
-        try:
-            lo = int(m.group(1).replace(',', '').split('.')[0])
-            hi = int(m.group(2).replace(',', '').split('.')[0])
-            if min(lo, hi) <= n <= max(lo, hi):
-                return True
-        except ValueError:
-            continue
-    return False
-
-
-def _close_to_any(n: int, candidates: set[str], tol: float) -> bool:
-    """candidates 중 하나라도 (1±tol) 배율 안에 들어오면 True.
-
-    n 은 정수, candidates 는 digits-only 문자열 집합.
-    """
-    for c in candidates:
-        try:
-            cn = int(c)
-        except ValueError:
-            continue
-        if cn == 0:
-            continue
-        if abs(n - cn) / cn <= tol:
-            return True
-    return False
-
-
-def _extract_content_tokens(text: str) -> set[str]:
-    """한국어 2자 이상 명사 + 영어 3자 이상 단어."""
-    return set(re.findall(r'[가-힣]{2,}|[a-zA-Z]{3,}', text))
-
-
-def _parse_number_with_unit(literal: str) -> tuple[str, str] | None:
-    """숫자 리터럴에서 (digits_only, unit) 분리. 단위 없으면 None."""
-    m = re.match(rf'([\d,.]+)\s*([{_UNIT_CHARS}])', literal)
-    if not m:
-        return None
-    digits = m.group(1).replace(',', '')
-    unit = m.group(2)
-    return (digits, unit)
-
-
-def _check_evidence_numeric_conflicts(evidence: list["EvidenceItem"]) -> list[str]:
-    """evidence 간 숫자 충돌 감지 (Phase 3.5b). evidence >= 2 일 때만 활성.
-
-    동일 단위, 다른 숫자 → weak flag. "이상/이하/초과/미만" 포함 시 skip.
-    bare number 는 비교 안 함 (조항 번호 등 false positive 방지).
-    """
-    if len(evidence) < 2:
-        return []
-
-    # 각 evidence 에서 단위 있는 숫자 + threshold 여부 추출
-    # {evidence_idx: [(digits, unit, has_threshold), ...]}
-    per_evidence: dict[int, list[tuple[str, str, bool]]] = {}
-    for idx, ev in enumerate(evidence):
-        nums = re.findall(
-            rf'\d[\d,.]*\s*[{_UNIT_CHARS}]\w{{0,4}}',
-            ev.span_text,
-        )
-        entries = []
-        for raw in nums:
-            parsed = _parse_number_with_unit(raw)
-            if not parsed:
-                continue
-            has_thr = bool(_THRESHOLD_SUFFIXES.search(raw))
-            entries.append((parsed[0], parsed[1], has_thr))
-        if entries:
-            per_evidence[idx] = entries
-
-    if len(per_evidence) < 2:
-        return []
-
-    # 단위별로 evidence 간 숫자 비교
-    # {unit: {digits: [evidence_idx, ...]}}
-    unit_map: dict[str, dict[str, list[int]]] = {}
-    for idx, entries in per_evidence.items():
-        for digits, unit, has_thr in entries:
-            if has_thr:
-                continue  # threshold 표현은 skip
-            if unit not in unit_map:
-                unit_map[unit] = {}
-            if digits not in unit_map[unit]:
-                unit_map[unit][digits] = []
-            if idx not in unit_map[unit][digits]:
-                unit_map[unit][digits].append(idx)
-
-    flags: list[str] = []
-    for unit, digits_map in unit_map.items():
-        distinct_values = list(digits_map.keys())
-        if len(distinct_values) >= 2:
-            # 가장 많이 등장하는 2개 비교
-            top2 = sorted(distinct_values, key=lambda d: len(digits_map[d]), reverse=True)[:2]
-            flags.append(
-                f"evidence_numeric_conflict:{top2[0]}{unit}_vs_{top2[1]}{unit}"
-            )
-
-    return flags
-
-
-def check(
-    query: str,
-    answer: str,
-    evidence: list[EvidenceItem],
-) -> GroundingResult:
-    """답변 vs evidence grounding 검증 + query intent alignment."""
-    strong: list[str] = []
-    weak: list[str] = []
-
-    if not answer or not evidence:
-        return GroundingResult([], [])
-
-    # ⚠ citation marker [n] 양측 제거 (대칭성 — Phase 3.5 B1)
-    evidence_text = re.sub(r'\[\d+\]', '', " ".join(e.span_text for e in evidence))
-
-    # ── Strong 1: fabricated number (unit-aware 3단계 — Phase 3.5 B1 fix1+fix3) ──
-    # Codex 지적 반영:
-    #   - fix1: range/tolerance/exact 모두 단위 일치 시에만 clear
-    #     (예: "150원" vs "100~200명" → flag 유지)
-    #   - fix3: 최대/최소 prefix 는 bound 의미 보존
-    #     (예: "최대 100명" + answer "100명" → flag 유지, "최대 100명" + answer "50명" → cleared)
-    answer_clean = re.sub(r'\[\d+\]', '', answer)
-    answer_corpus = _extract_numeric_corpus(answer_clean)
-    evidence_corpus = _extract_numeric_corpus(evidence_text)
-    ev_exact_by_unit = evidence_corpus["exact_by_unit"]
-    ev_ranges_by_unit = evidence_corpus["ranges_by_unit"]
-
-    # cleared 는 (unit, digits) 쌍 단위로 추적 — 단위 충돌 케이스 방어
-    cleared_pairs: set[tuple[str | None, str]] = set()
-
-    # Pass 1: 각 (unit, digits) 가 evidence 에서 정당화되는지 판정
-    for unit, digits_set in answer_corpus["exact_by_unit"].items():
-        for d in digits_set:
-            # 1) exact match — 같은 unit bucket 내에서만
-            if d in ev_exact_by_unit.get(unit, set()):
-                cleared_pairs.add((unit, d))
-                continue
-            # bare answer (unit=None) 는 evidence bare bucket 도 보조 매칭
-            if unit is None and d in ev_exact_by_unit.get(None, set()):
-                cleared_pairs.add((unit, d))
-                continue
-            try:
-                n = int(d)
-            except ValueError:
-                continue
-            # 2) range — same-unit 만 (bare answer 는 range clear 대상 아님)
-            if _within_unit_range(n, unit, ev_ranges_by_unit):
-                cleared_pairs.add((unit, d))
-                continue
-            # 3) ±1% tolerance — 단위가 양적(_TOLERANCE_UNITS) + 4자리+ + same-unit
-            if (
-                unit in _TOLERANCE_UNITS
-                and len(d) >= 4
-                and _close_to_unit_pool(n, unit, ev_exact_by_unit, tol=0.01)
-            ):
-                cleared_pairs.add((unit, d))
-                continue
-            # 식별자성 단위(_EXACT_ONLY_UNITS) 는 tolerance 패스 X.
-
-    # Pass 2: cleared 되지 않은 (unit, digits) 를 strong flag.
-    # 1자리 무시는 unit 이 식별자성(_EXACT_ONLY_UNITS: 년/월/일/조/항/호/회) 이 아닐 때만 적용.
-    # bare(None) 답변 숫자는 같은 digit 이 다른 unit 에서 cleared 됐으면 skip — 추출 부산물 방어.
-    # ⚠ 단위 cross-clear (예: "원" cleared → "명" 도 skip) 은 금지: Codex unit-mismatch 케이스가 깨짐.
-    unit_anchored_cleared: set[str] = {d for (u, d) in cleared_pairs if u is not None}
-    flagged_keys: set[tuple[str | None, str]] = set()
-    for unit, digits_set in answer_corpus["exact_by_unit"].items():
-        for d in digits_set:
-            if (unit, d) in cleared_pairs or (unit, d) in flagged_keys:
-                continue
-            # bare(None) 답변 숫자가 임의의 단위 bucket 에서 cleared 됐으면 duplicate 로 처리.
-            # 사례: "1,000명" → unit bucket "명" 에 1000 + bare bucket None 에 1000 (comma normalize 부산물).
-            # 이미 ("명", "1000") 가 cleared 라면 (None, "1000") 도 같은 사실을 가리키므로 skip.
-            if unit is None and d in unit_anchored_cleared:
-                continue
-            if len(d) < 2 and unit not in _EXACT_ONLY_UNITS:
-                continue
-            flagged_keys.add((unit, d))
-            # 사람이 읽기 좋게 "{digits}{unit}" 또는 bare 형태로 표기
-            label = f"{d}{unit}" if unit else d
-            strong.append(f"fabricated_number:{label}")
-
-    # ── Strong/Weak 2: query-answer intent alignment ──
-    query_content = _extract_content_tokens(query)
-    answer_content = _extract_content_tokens(answer)
-    if query_content:
-        missing_terms = query_content - answer_content
-        important_missing = [
-            t for t in missing_terms
-            if t not in GENERIC_TERMS and len(t) >= 2
-        ]
-        if important_missing:
-            strong.append(
-                f"intent_misalignment:{','.join(important_missing[:3])}"
-            )
-        elif len(missing_terms) > len(query_content) * 0.5:
-            weak.append(
-                f"intent_misalignment_generic:"
-                f"missing({','.join(list(missing_terms)[:5])})"
-            )
-
-    # ── Weak 1: uncited claim ──
-    sentences = re.split(r'(?<=[.!?。])\s+', answer)
-    for s in sentences:
-        if len(s.strip()) > 20 and not re.search(r'\[\d+\]', s):
-            weak.append(f"uncited_claim:{s[:40]}")
-
-    # ── Weak: evidence 간 숫자 충돌 (Phase 3.5b) ──
-    conflicts = _check_evidence_numeric_conflicts(evidence)
-    weak.extend(conflicts)
-
-    # ── Weak 2: token overlap ──
-    answer_tokens = _extract_content_tokens(answer)
-    evidence_tokens = _extract_content_tokens(evidence_text)
-    if answer_tokens:
-        overlap = len(answer_tokens & evidence_tokens) / len(answer_tokens)
-        if overlap < 0.4:
-            weak.append(f"low_overlap:{overlap:.2f}")
-
-    if strong or weak:
-        logger.info(
-            "grounding query=%r strong=%d weak=%d flags=%s",
-            query[:60],
-            len(strong),
-            len(weak),
-            ",".join(strong[:3] + weak[:3]),
-        )
-
-    return GroundingResult(strong, weak)
diff --git a/app/services/search/refusal_gate.py b/app/services/search/refusal_gate.py
deleted file mode 100644
index 60eff67..0000000
--- a/app/services/search/refusal_gate.py
+++ /dev/null
@@ -1,105 +0,0 @@
-"""Refusal gate — multi-signal fusion (Phase 3.5a).
-
-Score gate (deterministic) + classifier verdict (semantic, binary) 를 독립 평가 후 합성.
-Classifier 부재 시 3-tier conservative fallback.
-
-P1 실측 결과: exaone ternary 불안정 → binary (sufficient/insufficient) 로 축소.
-"full" vs "partial" 구분은 grounding check (intent alignment) 가 담당.
-"""
-
-from __future__ import annotations
-
-from dataclasses import dataclass, field
-from typing import TYPE_CHECKING, Literal
-
-from core.utils import setup_logger
-
-if TYPE_CHECKING:
-    from .classifier_service import ClassifierResult
-
-logger = setup_logger("refusal_gate")
-
-# Placeholder thresholds — Phase 3.5b 에서 실측 기반 tuning
-# AND 조건이라 false refusal 방어됨 (둘 다 만족해야 refuse)
-SCORE_MAX_REFUSE = 0.25
-SCORE_AGG_REFUSE = 0.70
-
-# Conservative fallback tiers (classifier 부재 시)
-CONSERVATIVE_WEAK = 0.35
-CONSERVATIVE_MID = 0.55
-
-
-@dataclass(slots=True)
-class RefusalDecision:
-    refused: bool
-    confidence_cap: Literal["high", "medium", "low"] | None  # None = no cap
-    rule_triggered: str | None  # 디버깅: 어느 signal 이 결정에 기여?
-
-
-def decide(
-    rerank_scores: list[float],
-    classifier: ClassifierResult | None,
-) -> RefusalDecision:
-    """Multi-signal fusion. Binary classifier verdict 기반.
-
-    Returns:
-        RefusalDecision. refused=True 이면 synthesis skip.
-        confidence_cap 은 synthesis 결과의 confidence 에 upper bound 적용.
-    """
-    max_score = max(rerank_scores) if rerank_scores else 0.0
-    agg_top3 = sum(sorted(rerank_scores, reverse=True)[:3])
-
-    score_gate_fails = (
-        max_score < SCORE_MAX_REFUSE and agg_top3 < SCORE_AGG_REFUSE
-    )
-
-    # ── Classifier 사용 가능 (정상 경로) ──
-    if classifier and classifier.verdict is not None:
-        if classifier.verdict == "insufficient":
-            # Evidence quality override: classifier 가 insufficient 라 해도
-            # evidence 가 충분히 좋으면 override (토론 8라운드 합의)
-            # (evidence quality 는 이 함수 밖에서 별도 체크 — caller 에서 처리)
-            logger.info(
-                "refusal gate: classifier=insufficient max=%.2f agg=%.2f",
-                max_score, agg_top3,
-            )
-            return RefusalDecision(
-                refused=True,
-                confidence_cap=None,
-                rule_triggered="classifier_insufficient",
-            )
-        if score_gate_fails:
-            logger.info(
-                "refusal gate: score_low max=%.2f agg=%.2f classifier=%s",
-                max_score, agg_top3, classifier.verdict,
-            )
-            return RefusalDecision(
-                refused=True,
-                confidence_cap=None,
-                rule_triggered="score_low",
-            )
-        # Classifier says sufficient → proceed
-        return RefusalDecision(
-            refused=False,
-            confidence_cap=None,
-            rule_triggered=None,
-        )
-
-    # ── Classifier 부재 → 3-tier conservative ──
-    if max_score < CONSERVATIVE_WEAK:
-        return RefusalDecision(
-            refused=True,
-            confidence_cap=None,
-            rule_triggered="conservative_refuse(no_classifier)",
-        )
-    if max_score < CONSERVATIVE_MID:
-        return RefusalDecision(
-            refused=False,
-            confidence_cap="low",
-            rule_triggered="conservative_low(no_classifier)",
-        )
-    return RefusalDecision(
-        refused=False,
-        confidence_cap="medium",
-        rule_triggered="conservative_medium(no_classifier)",
-    )
diff --git a/app/services/search/verifier_service.py b/app/services/search/verifier_service.py
deleted file mode 100644
index c2a8fe5..0000000
--- a/app/services/search/verifier_service.py
+++ /dev/null
@@ -1,196 +0,0 @@
-"""Exaone semantic verifier (Phase 3.5b).
-
-답변-근거 간 의미적 모순(contradiction) 감지. rule-based grounding_check 가 못 잡는
-미묘한 모순 포착. classifier 와 동일 패턴: circuit breaker + timeout + fail open.
-
-## Severity 3단계
-- strong: direct_negation (완전 모순) → re-gate 교차 자격
-- medium: numeric_conflict, intent_core_mismatch → confidence 하향 (누적 시 강제 low)
-- weak: nuance, unsupported_claim → 로깅 + mild confidence 하향
-
-## 핵심 원칙
-- **Verifier strong 단독 refuse 금지** — grounding strong 과 교차해야 refuse
-- **Timeout 3s** — 느리면 없는 게 낫다 (fail open)
-- MLX gate 사용 (Mac mini 26B endpoint — classifier/evidence 와 동일 gate 공유, 동시 race 방지)
-"""
-
-from __future__ import annotations
-
-import asyncio
-import os
-import time
-from dataclasses import dataclass, field
-from typing import TYPE_CHECKING, Literal
-
-from ai.client import AIClient, _load_prompt, parse_json_response
-from core.config import settings
-from core.utils import setup_logger
-from .llm_gate import Priority, acquire_mlx_gate
-
-if TYPE_CHECKING:
-    from .evidence_service import EvidenceItem
-
-logger = setup_logger("verifier")
-
-LLM_TIMEOUT_MS = 10000  # 2026-05-17 B-3: 3s 시 동시 부하 시 verifier 빈발 skip → grounding 약화. Mac mini 26B 가 verifier-style 짧은 LLM call 도 concurrent 호출 시 3s 초과 빈번 — 10s 로 raise
-CIRCUIT_THRESHOLD = 5
-CIRCUIT_RECOVERY_SEC = 60
-
-_failure_count = 0
-_circuit_open_until: float | None = None
-
-# Phase 3.5 B2: numeric_conflict severity promote 실험.
-# import time 평가 — env 변경 후 process restart 필수 (docker compose restart fastapi).
-# default=0 (off). production 적용은 B3 FP 검증 통과 후만.
-_NUMERIC_PROMOTE = os.getenv("VERIFIER_NUMERIC_PROMOTE", "0") == "1"
-
-# severity 매핑 (프롬프트 "critical"/"minor" → 코드 strong/medium/weak)
-# Tier 4 (B2): _NUMERIC_PROMOTE=1 일 때 numeric_conflict critical → strong 으로 격상.
-# minor 는 medium 유지 (FP 위험 분리).
-_SEVERITY_MAP: dict[str, dict[str, Literal["strong", "medium", "weak"]]] = {
-    "direct_negation": {"critical": "strong", "minor": "strong"},
-    "numeric_conflict": (
-        {"critical": "strong", "minor": "medium"} if _NUMERIC_PROMOTE
-        else {"critical": "medium", "minor": "medium"}
-    ),
-    "intent_core_mismatch": {"critical": "medium", "minor": "medium"},
-    "nuance": {"critical": "weak", "minor": "weak"},
-    "unsupported_claim": {"critical": "weak", "minor": "weak"},
-}
-
-
-@dataclass(slots=True)
-class Contradiction:
-    """개별 모순 발견."""
-    type: str  # direct_negation / numeric_conflict / intent_core_mismatch / nuance / unsupported_claim
-    severity: Literal["strong", "medium", "weak"]
-    claim: str
-    evidence_ref: str
-    explanation: str
-
-
-@dataclass(slots=True)
-class VerifierResult:
-    status: Literal["ok", "timeout", "error", "circuit_open", "skipped"]
-    contradictions: list[Contradiction]
-    elapsed_ms: float
-
-
-try:
-    VERIFIER_PROMPT = _load_prompt("verifier.txt")
-except FileNotFoundError:
-    VERIFIER_PROMPT = ""
-    logger.warning("verifier.txt not found — verifier will always skip")
-
-
-def _build_input(
-    answer: str,
-    evidence: list[EvidenceItem],
-) -> str:
-    """답변 + evidence spans → 프롬프트."""
-    spans = "\n\n".join(
-        f"[{e.n}] {(e.title or '').strip()}\n{e.span_text}"
-        for e in evidence
-    )
-    return (
-        VERIFIER_PROMPT
-        .replace("{answer}", answer)
-        .replace("{numbered_evidence}", spans)
-    )
-
-
-def _map_severity(ctype: str, raw_severity: str) -> Literal["strong", "medium", "weak"]:
-    """type + raw severity → 코드 severity 3단계."""
-    type_map = _SEVERITY_MAP.get(ctype, {"critical": "weak", "minor": "weak"})
-    return type_map.get(raw_severity, "weak")
-
-
-async def verify(
-    query: str,
-    answer: str,
-    evidence: list[EvidenceItem],
-) -> VerifierResult:
-    """답변-근거 semantic 검증. Parallel with grounding_check.
-
-    Returns:
-        VerifierResult. status "ok" 이 아니면 contradictions 빈 리스트 (fail open).
-    """
-    global _failure_count, _circuit_open_until
-    t_start = time.perf_counter()
-
-    if _circuit_open_until and time.time() < _circuit_open_until:
-        return VerifierResult("circuit_open", [], 0.0)
-
-    if not VERIFIER_PROMPT:
-        return VerifierResult("skipped", [], 0.0)
-
-    if not hasattr(settings.ai, "verifier") or settings.ai.verifier is None:
-        return VerifierResult("skipped", [], 0.0)
-
-    if not answer or not evidence:
-        return VerifierResult("skipped", [], 0.0)
-
-    prompt = _build_input(answer, evidence)
-    client = AIClient()
-    try:
-        async with acquire_mlx_gate(Priority.FOREGROUND):
-            async with asyncio.timeout(LLM_TIMEOUT_MS / 1000):
-                raw = await client.call_verifier(prompt)
-        _failure_count = 0
-    except asyncio.TimeoutError:
-        _failure_count += 1
-        if _failure_count >= CIRCUIT_THRESHOLD:
-            _circuit_open_until = time.time() + CIRCUIT_RECOVERY_SEC
-            logger.error(f"verifier circuit OPEN for {CIRCUIT_RECOVERY_SEC}s")
-        logger.warning("verifier timeout")
-        return VerifierResult(
-            "timeout", [],
-            (time.perf_counter() - t_start) * 1000,
-        )
-    except Exception as e:
-        _failure_count += 1
-        if _failure_count >= CIRCUIT_THRESHOLD:
-            _circuit_open_until = time.time() + CIRCUIT_RECOVERY_SEC
-            logger.error(f"verifier circuit OPEN for {CIRCUIT_RECOVERY_SEC}s")
-        logger.warning(f"verifier error: {e}")
-        return VerifierResult(
-            "error", [],
-            (time.perf_counter() - t_start) * 1000,
-        )
-    finally:
-        await client.close()
-
-    elapsed_ms = (time.perf_counter() - t_start) * 1000
-    parsed = parse_json_response(raw)
-    if not isinstance(parsed, dict):
-        logger.warning("verifier parse failed raw=%r", (raw or "")[:200])
-        return VerifierResult("error", [], elapsed_ms)
-
-    # contradiction 파싱
-    raw_items = parsed.get("contradictions") or []
-    if not isinstance(raw_items, list):
-        raw_items = []
-
-    results: list[Contradiction] = []
-    for item in raw_items[:5]:
-        if not isinstance(item, dict):
-            continue
-        ctype = item.get("type", "")
-        if ctype not in _SEVERITY_MAP:
-            ctype = "unsupported_claim"
-        raw_sev = item.get("severity", "minor")
-        severity = _map_severity(ctype, raw_sev)
-        claim = str(item.get("claim", ""))[:50]
-        ev_ref = str(item.get("evidence_ref", ""))[:50]
-        explanation = str(item.get("explanation", ""))[:30]
-        results.append(Contradiction(ctype, severity, claim, ev_ref, explanation))
-
-    logger.info(
-        "verifier ok query=%r contradictions=%d strong=%d medium=%d elapsed_ms=%.0f",
-        query[:60],
-        len(results),
-        sum(1 for c in results if c.severity == "strong"),
-        sum(1 for c in results if c.severity == "medium"),
-        elapsed_ms,
-    )
-    return VerifierResult("ok", results, elapsed_ms)
diff --git a/tests/api/test_search_ask_macbook_503.py b/tests/api/test_search_ask_macbook_503.py
deleted file mode 100644
index a6cdd19..0000000
--- a/tests/api/test_search_ask_macbook_503.py
+++ /dev/null
@@ -1,291 +0,0 @@
-"""PR-MacBook-RAG-Backend-1 정정 4 핵심 테스트.
-
-검증 invariant (synthesize 함수 레벨 — /ask wrapper 의 503 매핑은 search.py 의
-status="backend_unavailable" 분기로 1:1 deterministic):
-
-1. backend="qwen-macbook" + MacBook URL 죽은 포트
-   → synthesize() 가 SynthesisResult(status="backend_unavailable", ...) 반환
-   → Gemma backend 의 generate() 가 **단 1번도 호출되지 않음** (자동 fallback 부재)
-
-2. backend 미지정 (None)
-   → Gemma backend.generate() 호출, Qwen backend.generate() 호출 0
-   → 기존 호출자 (Hermes docsrv_ask / voice-memo-bot) 회귀 0
-
-3. backend="qwen-macbook" + MacBook 정상 응답
-   → status="completed" + answer 채워짐, Gemma backend 호출 0
-
-테스트 전략:
-- synthesize() 가 호출하는 backend dispatcher (services.llm.get_backend) 를
-  monkeypatch 해서 mock backend 주입.
-- Gemma backend 의 generate AsyncMock 호출 횟수를 추적.
-- 정정 4 의 핵심 가드: `gemma_backend.generate.assert_not_called()`
-"""
-
-from __future__ import annotations
-
-import asyncio
-import os
-import sys
-from dataclasses import dataclass
-from unittest.mock import AsyncMock
-
-import pytest
-
-sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "..", "app"))
-
-
-# ── 가짜 evidence (synthesize 의 no_evidence 분기 회피용 최소 객체) ─────────
-
-
-@dataclass
-class _FakeEvidence:
-    n: int = 1
-    doc_id: int = 100
-    chunk_id: int | None = 200
-    title: str | None = "fake doc"
-    span_text: str = "이것은 짧은 근거 텍스트입니다."
-    source: str = "llm"
-
-
-def _make_evidence():
-    return [_FakeEvidence()]
-
-
-# ── backend mock ───────────────────────────────────────────────────────────
-
-
-def _gemma_mock(content: str = "GEMMA_SHOULD_NEVER_BE_CALLED"):
-    m = AsyncMock()
-    m.name = "gemma-macmini"
-    m.generate = AsyncMock(return_value=content)
-    return m
-
-
-def _qwen_mock_success(content: str):
-    m = AsyncMock()
-    m.name = "qwen-macbook"
-    m.generate = AsyncMock(return_value=content)
-    return m
-
-
-def _qwen_mock_unavailable():
-    from services.llm import BackendUnavailable
-
-    m = AsyncMock()
-    m.name = "qwen-macbook"
-    m.generate = AsyncMock(
-        side_effect=BackendUnavailable("qwen-macbook", "ConnectError")
-    )
-    return m
-
-
-# ── 공통 fixture: synthesis_service 에 mock backend 주입 ───────────────────
-
-
-@pytest.fixture
-def patched_backends(monkeypatch):
-    """services.llm.get_backend 를 mock dispatcher 로 치환.
-
-    Returns (gemma_mock, qwen_mock, set_qwen_unavailable_fn).
-    """
-    from services.search import synthesis_service
-
-    gemma = _gemma_mock()
-    qwen_holder = {"backend": _qwen_mock_success(
-        '{"answer":"Qwen ok [1]","confidence":"high","refused":false}'
-    )}
-
-    def _fake_get_backend(name: str | None):
-        key = (name or "").strip().lower() or "gemma-macmini"
-        if key == "gemma-macmini":
-            return gemma
-        if key == "qwen-macbook":
-            return qwen_holder["backend"]
-        raise ValueError(f"unknown backend: {name!r}")
-
-    monkeypatch.setattr(synthesis_service, "get_backend", _fake_get_backend)
-    # synthesis_service 캐시 비움 (qwen vs gemma 캐시 분리 invariant)
-    synthesis_service._CACHE.clear()
-
-    def _swap_qwen_unavailable():
-        qwen_holder["backend"] = _qwen_mock_unavailable()
-
-    return gemma, qwen_holder, _swap_qwen_unavailable
-
-
-# ── 정정 4 핵심: backend=qwen-macbook + MacBook 비가용 → Gemma 호출 0 ─────
-
-
-def test_qwen_unavailable_yields_backend_unavailable_status_and_gemma_not_called(
-    patched_backends,
-):
-    """**정정 4 의 핵심 invariant**.
-
-    backend="qwen-macbook" 명시 + Qwen 호출이 BackendUnavailable 로 실패 →
-    synthesize() 는 status="backend_unavailable" 반환. Gemma backend 의
-    generate() 는 **단 한 번도 호출되지 않음** (silent fallback 금지).
-    """
-    from services.search.synthesis_service import synthesize
-
-    gemma, qwen_holder, swap_qwen_unavailable = patched_backends
-    swap_qwen_unavailable()
-    qwen = qwen_holder["backend"]
-
-    result = asyncio.run(
-        synthesize(
-            query="압력용기 최대허용응력은?",
-            evidence=_make_evidence(),
-            backend="qwen-macbook",
-        )
-    )
-
-    # 1. status
-    assert result.status == "backend_unavailable"
-    assert result.answer is None
-    assert result.confidence is None
-    assert result.refused is False
-
-    # 2. flag 에 backend 비가용 사유 기록
-    assert any(
-        f.startswith("backend_unavailable:qwen-macbook:") for f in result.hallucination_flags
-    ), f"expected backend_unavailable flag, got {result.hallucination_flags}"
-
-    # 3. ★ 핵심 가드 ★ — Gemma backend 자동 fallback 금지
-    gemma.generate.assert_not_called()
-
-    # 4. Qwen 은 1회만 호출 (재시도 없음)
-    assert qwen.generate.call_count == 1
-
-
-def test_qwen_unavailable_result_not_cached(patched_backends):
-    """비가용 결과는 캐시 X — 다음 호출이 다시 Qwen 시도해야 함."""
-    from services.search.synthesis_service import synthesize
-
-    gemma, qwen_holder, swap_qwen_unavailable = patched_backends
-    swap_qwen_unavailable()
-    qwen = qwen_holder["backend"]
-
-    asyncio.run(
-        synthesize(
-            query="동일 쿼리",
-            evidence=_make_evidence(),
-            backend="qwen-macbook",
-        )
-    )
-    asyncio.run(
-        synthesize(
-            query="동일 쿼리",
-            evidence=_make_evidence(),
-            backend="qwen-macbook",
-        )
-    )
-
-    # 두 번 모두 실제 호출 (캐시 적중 X) — Gemma 는 여전히 0
-    assert qwen.generate.call_count == 2
-    gemma.generate.assert_not_called()
-
-
-# ── 정정 4: backend 미지정 → 기존 Gemma path (회귀 0) ─────────────────────
-
-
-def test_default_backend_calls_gemma_not_qwen(patched_backends):
-    """backend 미지정 = 기본 Gemma. Qwen 호출 0."""
-    from services.search.synthesis_service import synthesize
-
-    gemma, qwen_holder, _ = patched_backends
-    qwen = qwen_holder["backend"]
-    gemma.generate.return_value = (
-        '{"answer":"Gemma 답변 [1]","confidence":"high","refused":false}'
-    )
-
-    result = asyncio.run(
-        synthesize(
-            query="기본 호출",
-            evidence=_make_evidence(),
-            backend=None,  # 명시 None = default
-        )
-    )
-
-    assert result.status == "completed"
-    assert result.answer is not None and "Gemma" in result.answer
-
-    # Qwen 은 호출 0
-    qwen.generate.assert_not_called()
-    # Gemma 는 1회
-    assert gemma.generate.call_count == 1
-
-
-# ── backend="qwen-macbook" + 정상 응답 ──────────────────────────────────────
-
-
-def test_qwen_success_does_not_call_gemma(patched_backends):
-    """Qwen 정상 응답 시 Gemma 는 호출되지 않음 (대칭 invariant)."""
-    from services.search.synthesis_service import synthesize
-
-    gemma, qwen_holder, _ = patched_backends
-    qwen = qwen_holder["backend"]
-
-    result = asyncio.run(
-        synthesize(
-            query="정상 호출",
-            evidence=_make_evidence(),
-            backend="qwen-macbook",
-        )
-    )
-
-    assert result.status == "completed"
-    assert result.answer is not None and "Qwen" in result.answer
-
-    # Gemma 는 0회
-    gemma.generate.assert_not_called()
-    # Qwen 은 1회
-    assert qwen.generate.call_count == 1
-
-
-# ── 캐시 분리 (qwen vs gemma 키 충돌 없음) ─────────────────────────────────
-
-
-def test_qwen_and_gemma_have_separate_caches(patched_backends):
-    """같은 query 라도 backend 다르면 캐시 분리 — Qwen 결과가 Gemma 호출 답으로 둔갑하지 않음."""
-    from services.search.synthesis_service import synthesize
-
-    gemma, qwen_holder, _ = patched_backends
-    qwen = qwen_holder["backend"]
-    gemma.generate.return_value = (
-        '{"answer":"GEMMA_ANSWER [1]","confidence":"high","refused":false}'
-    )
-    qwen.generate.return_value = (
-        '{"answer":"QWEN_ANSWER [1]","confidence":"high","refused":false}'
-    )
-
-    r_qwen_1 = asyncio.run(
-        synthesize(
-            query="같은 query",
-            evidence=_make_evidence(),
-            backend="qwen-macbook",
-        )
-    )
-    r_gemma_1 = asyncio.run(
-        synthesize(
-            query="같은 query",
-            evidence=_make_evidence(),
-            backend=None,
-        )
-    )
-    r_qwen_2 = asyncio.run(
-        synthesize(
-            query="같은 query",
-            evidence=_make_evidence(),
-            backend="qwen-macbook",
-        )
-    )
-
-    assert "QWEN_ANSWER" in (r_qwen_1.answer or "")
-    assert "GEMMA_ANSWER" in (r_gemma_1.answer or "")
-    # 두 번째 Qwen 호출은 캐시 적중 — 결과는 동일하지만 generate 추가 호출 X
-    assert "QWEN_ANSWER" in (r_qwen_2.answer or "")
-    assert r_qwen_2.cache_hit is True
-
-    # generate 호출 횟수: Qwen 1 (두번째는 캐시), Gemma 1
-    assert qwen.generate.call_count == 1
-    assert gemma.generate.call_count == 1
diff --git a/tests/api/test_search_ask_react_endpoint.py b/tests/api/test_search_ask_react_endpoint.py
deleted file mode 100644
index 380d70f..0000000
--- a/tests/api/test_search_ask_react_endpoint.py
+++ /dev/null
@@ -1,218 +0,0 @@
-"""PR-DocSrv-Ask-ToolCalling-ReAct-1: /api/search/ask/react endpoint integration.
-
-검증 항목 (G0-3 trace exposure + 정정 4 invariant):
-- backend unavailable → HTTP 503 + error_reason=macbook_unavailable
-  + ★ `run_search` mock 호출 횟수 == 0 (search 단계 진입 자체 차단)
-- 정상 응답 → 200 + final_answer + sources + debug_trace=null (default)
-- debug=true → debug_trace 채워짐
-- max rounds 도달 → iterations=2 + partial=false (final content 정상)
-
-endpoint 함수 (`api.search.ask_react`) 를 직접 호출하는 lightweight 패턴.
-TestClient 없이 FastAPI deps 를 MagicMock 으로 우회. (priority_gate / backend_dispatcher
-test 와 동일 service-layer 패턴.)
-"""
-
-from __future__ import annotations
-
-import asyncio
-import json
-import os
-import sys
-from unittest.mock import AsyncMock, MagicMock
-
-import pytest
-
-sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "..", "app"))
-
-
-# ── helpers ────────────────────────────────────────────────────────────────
-
-
-def _msg_with_tool_call(q: str, tc_id: str = "tc-1") -> dict:
-    return {
-        "role": "assistant",
-        "content": None,
-        "tool_calls": [
-            {
-                "id": tc_id,
-                "type": "function",
-                "function": {
-                    "name": "search",
-                    "arguments": json.dumps({"q": q}, ensure_ascii=False),
-                },
-            }
-        ],
-    }
-
-
-def _msg_with_content(text: str) -> dict:
-    return {"role": "assistant", "content": text, "tool_calls": None}
-
-
-def _fake_chunk(chunk_id: int, doc_id: int = 100):
-    m = MagicMock()
-    m.id = chunk_id
-    m.chunk_id = chunk_id
-    m.doc_id = doc_id
-    m.title = f"doc {doc_id}"
-    m.score = 0.9
-    m.snippet = f"snippet {chunk_id}"
-    m.text = None
-    return m
-
-
-def _fake_pr(chunks: list):
-    pr = MagicMock()
-    pr.results = chunks
-    return pr
-
-
-@pytest.fixture
-def patched_backend_and_search(monkeypatch):
-    """get_backend + run_search 둘 다 mock. backend 의 generate_with_tools 는
-    각 테스트가 side_effect 설정.
-
-    Returns: (backend_mock, run_search_mock, set_backend_unavailable_fn).
-    """
-    from services.llm.backends import BackendUnavailable, QwenMacBookBackend
-    from services.llm import backends as backends_mod
-    from services.search import react_loop
-
-    backend = MagicMock(spec=QwenMacBookBackend)
-    backend.name = "qwen-macbook"
-    backend.generate_with_tools = AsyncMock()
-
-    def _fake_get_backend(name):
-        # endpoint 가 qwen-macbook 만 호출하므로 단일 backend 반환
-        return backend
-
-    monkeypatch.setattr(backends_mod, "get_backend", _fake_get_backend)
-    # search.py 의 ask_react 안에서 `from services.llm.backends import ... get_backend`
-    # 로 import 하므로 module-level patch 만으로 충분 (지연 import 라 매번 fresh).
-
-    run_search_mock = AsyncMock(return_value=_fake_pr([_fake_chunk(1)]))
-    monkeypatch.setattr(react_loop, "run_search", run_search_mock)
-
-    def _make_unavailable():
-        backend.generate_with_tools.side_effect = BackendUnavailable(
-            "qwen-macbook", "ConnectError"
-        )
-
-    return backend, run_search_mock, _make_unavailable
-
-
-def _call_endpoint(payload):
-    """ask_react 를 직접 호출. user/session 은 MagicMock 으로 우회."""
-    from api.search import ask_react
-
-    user = MagicMock()
-    session = MagicMock()
-    return asyncio.run(ask_react(payload, user=user, session=session))
-
-
-# ── ★ 정정 4 invariant: backend unavailable → 503 + run_search 호출 0 ──────
-
-
-def test_qwen_unavailable_returns_503(patched_backend_and_search):
-    """backend BackendUnavailable → HTTP 503 + error_reason=macbook_unavailable."""
-    from api.search import AskReactRequest
-
-    backend, run_search_mock, make_unavailable = patched_backend_and_search
-    make_unavailable()
-
-    response = _call_endpoint(AskReactRequest(query="Q"))
-
-    # JSONResponse instance
-    assert response.status_code == 503
-    body = json.loads(response.body)
-    assert body["error_reason"] == "macbook_unavailable"
-    assert body["backend_used"] is None
-    assert body["backend_requested"] == "qwen-macbook"
-
-    # ★ run_search 호출 0 (search 진입 자체 차단)
-    assert run_search_mock.call_count == 0
-
-
-# ── 정상 200 + G0-3 default debug_trace=null ──────────────────────────────
-
-
-def test_successful_response_default_no_debug_trace(patched_backend_and_search):
-    """debug 미지정 (default false) → 200 + debug_trace == null."""
-    from api.search import AskReactRequest, AskReactResponse
-
-    backend, run_search_mock, _ = patched_backend_and_search
-    backend.generate_with_tools.side_effect = [
-        _msg_with_tool_call("q1"),
-        _msg_with_content("최종 답입니다"),
-    ]
-
-    response = _call_endpoint(AskReactRequest(query="Q"))
-
-    # Pydantic instance (FastAPI response_model 적용 전 raw return)
-    assert isinstance(response, AskReactResponse)
-    assert response.final_answer == "최종 답입니다"
-    assert response.iterations == 2
-    assert response.partial is False
-    assert response.debug_trace is None  # ★ G0-3
-    assert len(response.sources) == 1
-
-
-# ── G0-3: debug=true → debug_trace 채워짐 ──────────────────────────────────
-
-
-def test_debug_true_populates_trace(patched_backend_and_search):
-    from api.search import AskReactRequest
-
-    backend, run_search_mock, _ = patched_backend_and_search
-    backend.generate_with_tools.side_effect = [
-        _msg_with_content("바로 답"),
-    ]
-
-    response = _call_endpoint(AskReactRequest(query="Q", debug=True))
-
-    assert response.debug_trace is not None
-    assert isinstance(response.debug_trace, list)
-    assert len(response.debug_trace) >= 1
-
-
-# ── max rounds → final content 정상 → partial=false ──────────────────────
-
-
-def test_max_rounds_with_final_content(patched_backend_and_search):
-    from api.search import AskReactRequest
-
-    backend, run_search_mock, _ = patched_backend_and_search
-    backend.generate_with_tools.side_effect = [
-        _msg_with_tool_call("q1"),
-        _msg_with_tool_call("q2", tc_id="tc-2"),
-        _msg_with_content("정리된 최종 답"),
-    ]
-
-    response = _call_endpoint(AskReactRequest(query="Q"))
-
-    assert response.iterations == 2
-    assert response.partial is False
-    assert response.final_answer == "정리된 최종 답"
-    # LLM 호출 3회, search 2회 (G0-2 cap)
-    assert backend.generate_with_tools.call_count == 3
-    assert run_search_mock.call_count == 2
-
-
-# ── max rounds + final content 빈 string → partial=true ──────────────────
-
-
-def test_max_rounds_with_empty_final_partial(patched_backend_and_search):
-    from api.search import AskReactRequest
-
-    backend, run_search_mock, _ = patched_backend_and_search
-    backend.generate_with_tools.side_effect = [
-        _msg_with_tool_call("q1"),
-        _msg_with_tool_call("q2", tc_id="tc-2"),
-        _msg_with_content(""),
-    ]
-
-    response = _call_endpoint(AskReactRequest(query="Q"))
-
-    assert response.iterations == 2
-    assert response.partial is True
-    assert response.final_answer == ""
diff --git a/tests/test_ask_eval_auth.py b/tests/test_ask_eval_auth.py
deleted file mode 100644
index 0492d21..0000000
--- a/tests/test_ask_eval_auth.py
+++ /dev/null
@@ -1,92 +0,0 @@
-"""Phase 3.5 fix2: /ask 의 X-Source / X-Eval-Case-Id trust boundary.
-
-`_resolve_eval_identity()` 단위 테스트.
-- token 없음/틀림 + X-Source=eval → source='document_server', eval_case_id=None
-- token 일치 + X-Source=eval + X-Eval-Case-Id=case_xxx → ('eval', 'case_xxx')
-- token 틀림 + X-Eval-Case-Id 만 (X-Source 미지정) → eval_case_id=None
-- 일반 호출 (X-Source=ui_search, no eval headers) → ('ui_search', None)
-- env 미설정 (eval_runner_token='') 시 모든 eval claim 거부
-"""
-
-from __future__ import annotations
-
-import os
-import sys
-
-sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "app"))
-
-import pytest
-
-
-@pytest.fixture
-def resolve_with_token(monkeypatch):
-    """settings.eval_runner_token 을 monkey-patch 해서 _resolve_eval_identity 테스트."""
-    def _make(token: str):
-        from core import config as cfg_mod
-        from api import search as search_mod
-        # 두 모듈 모두에서 settings 객체 참조하므로 직접 attr 변경
-        monkeypatch.setattr(search_mod.settings, "eval_runner_token", token)
-        return search_mod._resolve_eval_identity
-    return _make
-
-
-def test_no_token_no_eval_headers_default(resolve_with_token):
-    """일반 호출 — eval 헤더 없음, source 기본값."""
-    resolve = resolve_with_token("secret123")
-    assert resolve(None, None, None) == ("document_server", None)
-
-
-def test_normal_source_with_token(resolve_with_token):
-    """ui_search 호출 — eval 클레임 아님이라 token 무관."""
-    resolve = resolve_with_token("secret123")
-    assert resolve("ui_search", None, None) == ("ui_search", None)
-
-
-def test_eval_claim_no_token_rejected(resolve_with_token):
-    """X-Source=eval 인데 token 없음 → 거부, source='document_server'."""
-    resolve = resolve_with_token("secret123")
-    assert resolve("eval", "case_001", None) == ("document_server", None)
-
-
-def test_eval_claim_wrong_token_rejected(resolve_with_token):
-    """token 틀림 → 거부."""
-    resolve = resolve_with_token("secret123")
-    assert resolve("eval", "case_001", "wrong_token") == ("document_server", None)
-
-
-def test_eval_claim_correct_token_accepted(resolve_with_token):
-    """token 일치 → 'eval' source + case_id 적재."""
-    resolve = resolve_with_token("secret123")
-    assert resolve("eval", "case_001", "secret123") == ("eval", "case_001")
-
-
-def test_eval_case_id_only_no_source_no_token(resolve_with_token):
-    """X-Eval-Case-Id 만 있고 token 없음 → 거부, case_id=None."""
-    resolve = resolve_with_token("secret123")
-    assert resolve(None, "case_001", None) == ("document_server", None)
-
-
-def test_eval_case_id_only_wrong_token(resolve_with_token):
-    """X-Eval-Case-Id 만 + token 틀림 → 거부."""
-    resolve = resolve_with_token("secret123")
-    assert resolve(None, "case_001", "wrong") == ("document_server", None)
-
-
-def test_env_unset_rejects_even_correct_format(resolve_with_token):
-    """settings.eval_runner_token='' 인 환경 → 모든 eval 클레임 거부."""
-    resolve = resolve_with_token("")
-    # token 헤더가 와도 server side 가 비어있으면 거부 (constant-time False)
-    assert resolve("eval", "case_001", "") == ("document_server", None)
-    assert resolve("eval", "case_001", "anything") == ("document_server", None)
-
-
-def test_non_eval_source_forces_case_id_none(resolve_with_token):
-    """X-Source=ui_detail + X-Eval-Case-Id (실수로 같이 보냄) → case_id=None.
-
-    eval claim 아님 (source != 'eval' 이고 case_id 가 fallback 으로 eval claim 트리거)
-    이지만 source claim 이 명시적으로 non-eval 이라 token 검증 후 case_id None.
-    """
-    resolve = resolve_with_token("secret123")
-    # case_id 가 있으면 eval claim 으로 처리됨 → token 없으면 거부 → ('ui_detail' 클레임,
-    # 하지만 거부 분기에서 claimed_source != 'eval' 이라 그대로 'ui_detail' 반환, case_id=None)
-    assert resolve("ui_detail", "case_001", None) == ("ui_detail", None)
diff --git a/tests/test_grounding_fabricated_number.py b/tests/test_grounding_fabricated_number.py
deleted file mode 100644
index ae76b6d..0000000
--- a/tests/test_grounding_fabricated_number.py
+++ /dev/null
@@ -1,188 +0,0 @@
-"""Phase 3.5 B1 (fix1+fix3): unit-aware fabricated_number + bound semantics.
-
-기준:
-- 단위 일치 시에만 exact/range/tolerance clear (fix1: Codex unit-mismatch regression 방지)
-- 약/대략/거의/얼추 만 approx prefix strip; 최대/최소 는 bound operator 로 보존 (fix3)
-- tolerance 는 양적 단위(_TOLERANCE_UNITS) + 4자리+ 만; 식별자성(_EXACT_ONLY_UNITS) 은 strict
-"""
-
-from __future__ import annotations
-
-import os
-import sys
-
-# tests/ → 프로젝트 루트 → app/
-sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "app"))
-
-import pytest
-
-from services.search.evidence_service import EvidenceItem
-from services.search.grounding_check import check
-
-
-def _ev(text: str, n: int = 1) -> EvidenceItem:
-    return EvidenceItem(
-        n=n,
-        chunk_id=None,
-        doc_id=100 + n,
-        title=f"doc{n}",
-        section_title=None,
-        span_text=text,
-        relevance=0.9,
-        rerank_score=0.85,
-        full_snippet=text,
-        source="llm",
-    )
-
-
-def _has_fabricated(result, sub: str | None = None) -> bool:
-    for f in result.strong_flags:
-        if not f.startswith("fabricated_number:"):
-            continue
-        if sub is None or sub in f:
-            return True
-    return False
-
-
-# ─── 콤마/prefix/range/단위 동의어/citation (기존 17 케이스) ──────
-
-
-def test_comma_thousand_match():
-    r = check("질문", "총 1,000명 [1]", [_ev("총원은 1000명입니다.")])
-    assert not _has_fabricated(r, "1000")
-
-
-def test_comma_thousand_reverse():
-    r = check("질문", "총 1000명 [1]", [_ev("총원은 1,000명입니다.")])
-    assert not _has_fabricated(r)
-
-
-def test_approx_prefix_in_answer():
-    r = check("질문", "약 100명이 참여 [1]", [_ev("100명이 참여")])
-    assert not _has_fabricated(r)
-
-
-def test_approx_prefix_in_evidence():
-    r = check("질문", "100명이 참여 [1]", [_ev("약 100명이 참여")])
-    assert not _has_fabricated(r)
-
-
-def test_range_inner_value_passes():
-    r = check("질문", "약 150명 [1]", [_ev("100~200명 사이 추정")])
-    assert not _has_fabricated(r, "150")
-
-
-def test_range_outer_value_flagged():
-    r = check("질문", "300명 [1]", [_ev("100~200명 사이 추정")])
-    assert _has_fabricated(r, "300")
-
-
-def test_unit_synonym_in_to_myeong():
-    r = check("질문", "총 50인이 모임 [1]", [_ev("총 50명이 모임.")])
-    assert not _has_fabricated(r)
-
-
-def test_unit_synonym_percent_to_pct():
-    r = check("질문", "비율 30퍼센트 [1]", [_ev("비율 30%이다.")])
-    assert not _has_fabricated(r)
-
-
-def test_citation_marker_both_sides():
-    """bug fix: evidence 측 [n] 미제거로 디지트 합쳐지던 케이스."""
-    r = check("질문", "가격 [1] 5,000원", [_ev("[2] 5,000원이 정확")])
-    assert not _has_fabricated(r)
-
-
-def test_genuine_fabricated_number():
-    r = check("질문", "결과 777명 [1]", [_ev("500명, 300명을 받음.")])
-    assert _has_fabricated(r, "777")
-
-
-def test_amount_4digit_tolerance_passes():
-    r = check("질문", "9,990원 [1]", [_ev("10,000원입니다.")])
-    assert not _has_fabricated(r)
-
-
-def test_year_no_tolerance_flagged():
-    r = check("질문", "2024년 [1]", [_ev("2026년에 발효")])
-    assert _has_fabricated(r, "2024")
-
-
-def test_article_no_tolerance_flagged():
-    r = check("질문", "제5조에 명시 [1]", [_ev("제6조에 따라")])
-    assert _has_fabricated(r)
-
-
-def test_count_no_tolerance_flagged():
-    r = check("질문", "총 3회 위반 [1]", [_ev("총 4회 적발")])
-    assert _has_fabricated(r)
-
-
-def test_three_digit_strict():
-    r = check("질문", "총 15개 [1]", [_ev("총 10개")])
-    assert _has_fabricated(r, "15")
-
-
-def test_single_digit_ignored():
-    """1자리 + 양적 단위 → 무시 (오탐 방지)."""
-    r = check("질문", "총 3개 발생 [1]", [_ev("관련 통계 별도")])
-    assert not _has_fabricated(r, "3개")
-
-
-def test_range_korean_butter_separator():
-    r = check("질문", "약 150명 [1]", [_ev("100부터 200명까지 대상.")])
-    assert not _has_fabricated(r, "150")
-
-
-# ─── fix1: unit-mismatch (Codex no-ship) ──────────────────
-
-
-def test_won_vs_myeong_range_flagged():
-    """answer '150원' vs evidence '100~200명' → 단위 불일치, flag 유지."""
-    r = check("질문", "약 150원이 든다 [1]", [_ev("대상은 100~200명")])
-    assert _has_fabricated(r, "150")
-
-
-def test_won_vs_myeong_tolerance_flagged():
-    """answer '9,990원' vs evidence '10,000명' → tolerance pool 단위 다름, flag 유지."""
-    r = check("질문", "9,990원 [1]", [_ev("10,000명입니다.")])
-    assert _has_fabricated(r, "9990")
-
-
-def test_pct_vs_myeong_range_flagged():
-    """answer '15%' vs evidence '10~20명' → 단위 불일치, flag 유지."""
-    r = check("질문", "약 15% [1]", [_ev("대상 10~20명")])
-    assert _has_fabricated(r, "15")
-
-
-# ─── fix3: 최대/최소 bound semantics ───────────────────────
-
-
-def test_choedae_exact_boundary_flagged():
-    """evidence '최대 100명' + answer '100명' → 경계값 자체는 cleared 아님."""
-    r = check("질문", "100명이다 [1]", [_ev("최대 100명까지 가능")])
-    assert _has_fabricated(r, "100")
-
-
-def test_choeso_exact_boundary_flagged():
-    """evidence '최소 100명' + answer '100명' → 경계값 자체는 cleared 아님."""
-    r = check("질문", "100명이다 [1]", [_ev("최소 100명 이상 필요")])
-    assert _has_fabricated(r, "100")
-
-
-def test_choedae_inner_value_passes():
-    """evidence '최대 100명' + answer '50명' → bound 안, cleared."""
-    r = check("질문", "50명이다 [1]", [_ev("최대 100명까지 가능")])
-    assert not _has_fabricated(r, "50")
-
-
-def test_choeso_above_value_passes():
-    """evidence '최소 100명' + answer '150명' → bound 안, cleared."""
-    r = check("질문", "150명이다 [1]", [_ev("최소 100명 이상 필요")])
-    assert not _has_fabricated(r, "150")
-
-
-def test_choedae_outer_value_flagged():
-    """evidence '최대 100명' + answer '200명' → bound 밖, flag."""
-    r = check("질문", "200명이다 [1]", [_ev("최대 100명까지 가능")])
-    assert _has_fabricated(r, "200")
diff --git a/tests/test_synthesis_failure_regate.py b/tests/test_synthesis_failure_regate.py
deleted file mode 100644
index f06577f..0000000
--- a/tests/test_synthesis_failure_regate.py
+++ /dev/null
@@ -1,123 +0,0 @@
-"""Phase 3.5 fix3: re-gate Tier 0 — synthesis 자체 실패 처리.
-
-`_detect_synthesis_failure()` 단위 테스트.
-
-기존 버그:
-  synthesis LLM self-refuse (`sr.refused=True, status="completed"`) 또는
-  timeout/parse_failed/llm_error 시 grounding/verifier flag 0건 → re-gate else clean
-  분기로 빠져 `completeness="full"` 초기값이 남아 `full + refused=True` 모순.
-  baseline v1-400char 에서 24/223 (10.8%) 해당.
-
-Tier 0 판정:
-  - LLM self-refuse (completed + refused) → "synthesis_self_refuse"
-  - mechanical fail (timeout/parse_failed/llm_error) → "synthesis_failed({status})"
-  - answer 공백 → "synthesis_failed({status})"
-  - 유효 답변 → None (기존 tier 1~7 경로)
-"""
-
-from __future__ import annotations
-
-import os
-import sys
-
-sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "app"))
-
-from api.search import _detect_synthesis_failure
-from services.search.synthesis_service import SynthesisResult
-
-
-def _sr(
-    status: str = "completed",
-    answer: str | None = "ok",
-    refused: bool = False,
-    refuse_reason: str | None = None,
-) -> SynthesisResult:
-    return SynthesisResult(
-        status=status,  # type: ignore[arg-type]
-        answer=answer,
-        used_citations=[],
-        confidence="low",
-        refused=refused,
-        refuse_reason=refuse_reason,
-        elapsed_ms=100.0,
-        cache_hit=False,
-    )
-
-
-# ─── self-refuse 케이스 ──────────────────────────────────
-
-
-def test_llm_self_refuse_completed():
-    """LLM 이 JSON 에 refused=true 반환 → synthesis_self_refuse."""
-    sr = _sr(status="completed", answer=None, refused=True, refuse_reason="범위 밖")
-    assert _detect_synthesis_failure(sr) == "synthesis_self_refuse"
-
-
-def test_llm_self_refuse_with_answer_still_refused():
-    """refused=True 면 answer 있어도 Tier 0 처리 (일관성)."""
-    sr = _sr(status="completed", answer="왜 답변함", refused=True)
-    assert _detect_synthesis_failure(sr) == "synthesis_self_refuse"
-
-
-# ─── mechanical failure 케이스 ──────────────────────────
-
-
-def test_timeout():
-    sr = _sr(status="timeout", answer=None, refused=False)
-    assert _detect_synthesis_failure(sr) == "synthesis_failed(timeout)"
-
-
-def test_parse_failed():
-    sr = _sr(status="parse_failed", answer=None, refused=False)
-    assert _detect_synthesis_failure(sr) == "synthesis_failed(parse_failed)"
-
-
-def test_llm_error():
-    sr = _sr(status="llm_error", answer=None, refused=False)
-    assert _detect_synthesis_failure(sr) == "synthesis_failed(llm_error)"
-
-
-def test_refused_with_mechanical_fail_propagates_status():
-    """refused=True + status!=completed → synthesis_failed({status}) 형식."""
-    sr = _sr(status="timeout", answer=None, refused=True)
-    assert _detect_synthesis_failure(sr) == "synthesis_failed(timeout)"
-
-
-# ─── empty answer 케이스 ───────────────────────────────
-
-
-def test_empty_answer_completed():
-    """status=completed 인데 answer 공백 → synthesis_failed(completed)."""
-    sr = _sr(status="completed", answer="", refused=False)
-    assert _detect_synthesis_failure(sr) == "synthesis_failed(completed)"
-
-
-def test_whitespace_only_answer():
-    """공백/탭/개행만 있어도 empty 로 간주."""
-    sr = _sr(status="completed", answer="   \n\t  ", refused=False)
-    assert _detect_synthesis_failure(sr) == "synthesis_failed(completed)"
-
-
-def test_none_answer_completed():
-    """answer=None + status=completed → failed."""
-    sr = _sr(status="completed", answer=None, refused=False)
-    assert _detect_synthesis_failure(sr) == "synthesis_failed(completed)"
-
-
-# ─── 유효 답변 케이스 (None 반환) ──────────────────────
-
-
-def test_valid_answer_returns_none():
-    """status=completed + answer 있고 refused=False → Tier 0 통과 (None)."""
-    sr = _sr(status="completed", answer="교육 시간은 매년 6시간 이상이다 [1].", refused=False)
-    assert _detect_synthesis_failure(sr) is None
-
-
-def test_skipped_status_with_answer_passes():
-    """status=skipped 는 Tier 0 대상 아님 — 초기 refusal gate 에서 이미 early-return 처리됨.
-
-    (skipped 는 여기까지 도달하지 않는다는 전제. 만약 도달하더라도 refused 가 True 일 것.)
-    """
-    sr = _sr(status="skipped", answer="abc", refused=False)
-    # 이 경우 Tier 0 미발동 (answer 있고 refused 아님) — 정상 경로로 나감.
-    assert _detect_synthesis_failure(sr) is None
diff --git a/tests/test_verifier_numeric_promote.py b/tests/test_verifier_numeric_promote.py
deleted file mode 100644
index e7c8443..0000000
--- a/tests/test_verifier_numeric_promote.py
+++ /dev/null
@@ -1,58 +0,0 @@
-"""Phase 3.5 B2: verifier _SEVERITY_MAP env flag 테스트.
-
-VERIFIER_NUMERIC_PROMOTE 환경변수에 따른 _SEVERITY_MAP 변화 검증.
-모듈은 import time 에 env 평가하므로 reload 필요.
-"""
-
-from __future__ import annotations
-
-import importlib
-import os
-import sys
-
-# tests/ → 프로젝트 루트 → app/
-sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "app"))
-
-import pytest
-
-
-def _reload_verifier(monkeypatch, value: str | None):
-    """env 설정 후 verifier_service 를 reload 하여 _SEVERITY_MAP 재평가."""
-    if value is None:
-        monkeypatch.delenv("VERIFIER_NUMERIC_PROMOTE", raising=False)
-    else:
-        monkeypatch.setenv("VERIFIER_NUMERIC_PROMOTE", value)
-    from services.search import verifier_service
-    importlib.reload(verifier_service)
-    return verifier_service
-
-
-def test_severity_map_off_default(monkeypatch):
-    """env 미설정 → numeric_conflict critical 은 medium (기존 동작)."""
-    vs = _reload_verifier(monkeypatch, None)
-    assert vs._SEVERITY_MAP["numeric_conflict"]["critical"] == "medium"
-    assert vs._SEVERITY_MAP["numeric_conflict"]["minor"] == "medium"
-    assert vs._NUMERIC_PROMOTE is False
-
-
-def test_severity_map_on_critical_promoted(monkeypatch):
-    """VERIFIER_NUMERIC_PROMOTE=1 → critical 만 strong, minor 는 medium 유지."""
-    vs = _reload_verifier(monkeypatch, "1")
-    assert vs._SEVERITY_MAP["numeric_conflict"]["critical"] == "strong"
-    assert vs._SEVERITY_MAP["numeric_conflict"]["minor"] == "medium"
-    assert vs._NUMERIC_PROMOTE is True
-
-
-def test_severity_map_off_explicit_zero(monkeypatch):
-    """VERIFIER_NUMERIC_PROMOTE=0 명시 → off (default 와 동일)."""
-    vs = _reload_verifier(monkeypatch, "0")
-    assert vs._SEVERITY_MAP["numeric_conflict"]["critical"] == "medium"
-    assert vs._NUMERIC_PROMOTE is False
-
-
-def test_direct_negation_invariant(monkeypatch):
-    """direct_negation 은 env 무관 항상 strong (불변 — 안전장치)."""
-    for value in [None, "0", "1"]:
-        vs = _reload_verifier(monkeypatch, value)
-        assert vs._SEVERITY_MAP["direct_negation"]["critical"] == "strong"
-        assert vs._SEVERITY_MAP["direct_negation"]["minor"] == "strong"