feat(scoring): 六维评分行为锚点标尺(中/英同标尺,试点)

backend/app/services/llm/prompts/analysis.py

@@ -11,6 +11,7 @@
 from app.services.scoring_caliber import (
     curation_threshold_text,
     curation_weights_text,
+    dimension_anchor_section,
     risk_penalty_text,
 )
 
@@ -60,6 +61,10 @@
   "title_suggestions": [""]
 }}
 
+"""
+    + dimension_anchor_section()
+    + """
+
 精选分（curation_score）评判标准：
 - ≥80：重大发布/独家/强实用性工具/高传播力事件
 - 70-79：扎实的产品更新/行业动态/有价值教程
@@ -128,6 +133,10 @@
   "title_suggestions": [""]
 }}
 
+"""
+    + dimension_anchor_section("en")
+    + """
+
 Curation score (curation_score) guidelines:
 - ≥80: major release / exclusive / highly actionable tool / high-spread event
 - 70-79: solid product update / industry development / valuable tutorial

backend/app/services/scoring_caliber.py

@@ -115,3 +115,110 @@ def risk_penalty_text(lang: str = "zh") -> str:
     if lang == "en":
         return f"risk>{LLM_RISK_PENALTY_TRIGGER} deducts {LLM_RISK_PENALTY_POINTS} points"
     return f"风险分>{LLM_RISK_PENALTY_TRIGGER}则扣{LLM_RISK_PENALTY_POINTS}分"
+
+
+# ── 六维评分行为锚点(#114,C3 试点) ────────────────────────────────────
+# scores 块六维从「一句名称提示」升级为四档可判定行为描述,让打分跨样本、
+# 跨语种可比:同一个 85 分在任何一天意味着同一件事。每档写「可判定的行为
+# 特征」而非抽象程度词;中文为口径原文,英文为同一标尺译文。
+# 论文版 prompt 是否复用属独立决策(#114),首期只挂中/英两套。
+ANCHOR_BANDS: tuple[tuple[int, int], ...] = ((0, 30), (40, 60), (70, 85), (90, 100))
+
+DIMENSION_ANCHORS_ZH: dict[str, list[str]] = {
+    "quality_score": [
+        "纯转发、寒暄或情绪宣泄，信息量趋近于零，或逻辑断裂看不出主线",
+        "有明确观点且能自圆其说，但缺少数据、案例或可验证细节",
+        "论证完整，带数据/案例/步骤等可验证细节，主线清晰",
+        "一手深料：独家数据、完整方法论、系统实测或深度复盘",
+    ],
+    "hot_score": [
+        "无讨论量，发布于冷门角落，看不到互动信号",
+        "小圈子在讨论（评论数十量级），尚未出圈",
+        "多个社区或账号在转述，评论区活跃，趋势榜可见",
+        "刷屏级：热搜头部、多平台同时爆发、大V密集跟进",
+    ],
+    "freshness_score": [
+        "数月前的旧闻，或已被充分报道的旧信息",
+        "一周内的动态，或旧话题的新进展",
+        "24-72小时内的新发布或首发报道",
+        "独家首发、官宣/泄露后数小时、或当天突发",
+    ],
+    "creator_score": [
+        "与目标领域无关，或没有可延展的选题空间",
+        "领域相关可作背景素材，但角度已被写烂",
+        "能直接展开成一篇内容：角度明确、读者清晰、有差异化空间",
+        "选题富矿：可系列化、多角度深挖，或正处话题空白期",
+    ],
+    "viral_score": [
+        "话题封闭，只有当事人与极少数人关心",
+        "有一定共鸣点，但表达门槛高，难以通俗转述",
+        "有冲突、反常识或实用钩子，一句话能讲清，易被转发",
+        "踩中大众情绪点或强实用干货，天然适合标题与二创",
+    ],
+    "risk_score": [
+        "常规技术/产品/资讯内容，无可预见争议",
+        "轻度敏感：涉争议人物、营销夸大或立场分歧，需注意措辞",
+        "高风险：未证实的重大指控、医疗/投资建议类、平台限流领域",
+        "红线：违法、色情、歧视、谣言等，不建议触碰",
+    ],
+}
+
+DIMENSION_ANCHORS_EN: dict[str, list[str]] = {
+    "quality_score": [
+        "pure repost, small talk, or venting with near-zero information, or incoherent logic",
+        "has a clear, self-consistent viewpoint but lacks data, cases, or verifiable detail",
+        "complete argument with data/cases/steps and verifiable detail, clear throughline",
+        "first-hand depth: exclusive data, full methodology, systematic hands-on testing, or deep retrospective",
+    ],
+    "hot_score": [
+        "no discussion, posted in an obscure corner, no engagement signals",
+        "niche community discussing it (dozens of comments), not yet breaking out",
+        "multiple communities/accounts relaying it, active comments, visible on trend boards",
+        "explosive: top of trending, erupting across platforms, heavy influencer pickup",
+    ],
+    "freshness_score": [
+        "months-old news, or information already widely covered",
+        "a development within the past week, or a new twist on an old topic",
+        "new release or first report within 24-72 hours",
+        "exclusive debut, hours after announcement/leak, or same-day breaking",
+    ],
+    "creator_score": [
+        "irrelevant to the target domain, or no angle to build on",
+        "domain-relevant background material, but the angle is worn out",
+        "directly expandable into a piece: clear angle, clear audience, room to differentiate",
+        "topic goldmine: series-worthy, multi-angle, or filling a topical blank",
+    ],
+    "viral_score": [
+        "closed topic, only stakeholders and a handful of people care",
+        "some resonance, but hard to convey in plain words",
+        "conflict/counter-intuitive/practical hook, explainable in one sentence, easy to share",
+        "mass emotional appeal or strong utility, naturally title- and remix-friendly",
+    ],
+    "risk_score": [
+        "routine tech/product/news content, no foreseeable controversy",
+        "mildly sensitive: controversial figures, marketing hype, or divisive stances",
+        "high risk: unverified serious allegations, medical/investment advice, platform-throttled areas",
+        "red line: illegal, pornographic, discriminatory, or rumor content",
+    ],
+}
+
+
+def dimension_anchor_lines(lang: str = "zh") -> list[str]:
+    """六维行为锚点行(每维一行,四档以分号分隔),数值与文案同出本模块。"""
+    table = DIMENSION_ANCHORS_EN if lang == "en" else DIMENSION_ANCHORS_ZH
+    sep = "; " if lang == "en" else "；"
+    colon = ": " if lang == "en" else "："
+    lines: list[str] = []
+    for dim, descriptions in table.items():
+        parts = [f"{lo}-{hi}={descriptions[i]}" for i, (lo, hi) in enumerate(ANCHOR_BANDS)]
+        lines.append(f"- {dim}{colon}" + sep.join(parts))
+    return lines
+
+
+def dimension_anchor_section(lang: str = "zh") -> str:
+    """prompt 的完整锚点段落(标题行 + 六行锚点),中/英共用同一标尺。"""
+    if lang == "en":
+        header = "Six-dimension scoring anchors (score by band; interpolate between bands):"
+    else:
+        header = "六维评分锚点（按档给分，档间取值按程度插值）："
+    return header + "\n" + "\n".join(dimension_anchor_lines(lang))