refactor(search)!: /ask 고아 service·테스트·프롬프트 정리 (검색 단일화 Phase 2)

/ask 삭제로 0-consumer 된 자산 제거(3-gate 실증): search.py /ask 섹션(Citation/ConfirmedItem/AskDebug/AskResponse 모델 + 헬퍼 + _resolve_eval_identity) + 죽은 import 13개. service 4(classifier/verifier/refusal_gate/grounding_check). AIClient.call_classifier/call_verifier(고아). 프롬프트 2(classifier/verifier.txt). broken test 6. evidence/synthesis 는 공유(documents.py 등)라 유지. 실 pyflakes 클린(이전 세션 pyflakes 미설치로 검증 누락 → 설치 후 실검증).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

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:

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
-
-

app/prompts/classifier.txt

@@ -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}]

app/prompts/verifier.txt

@@ -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}

app/services/search/classifier_service.py

@@ -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)

app/services/search/grounding_check.py

@@ -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)

app/services/search/refusal_gate.py

@@ -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)",
-    )

app/services/search/verifier_service.py

@@ -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)