git:20260705.ef74a22 to git:20260706.6c09d25

111 added, 111 removed. Audit A to A.

What this update changes in the checks

No check changed: every fixed check gives the same answer before and after this update.

A tripped check is a label, not a verdict: read the matched text and the lines below. This week's updates that newly trip a check.

---
name: private-company-research
category: analysis
description: "Deep research framework for pre-IPO / private companies (Ant Group, SpaceX, Stripe, ByteDance...). Six analyst lenses — business model, financial forensics, competitive landscape, risk & governance, tech & IP, alternative-data signals — run in parallel via run_swarm, then cross-validated for signal consistency before any verdict. Built around the core challenge of private-company work: information is scarce, so every data point carries a confidence label (high / medium / low), inference is shown separately from fact, and 'I don't know' is a valid output. Outputs a fair-value range, exit-path analysis, and an information-gap map. Use for any unlisted company where you need to judge what the business is actually worth."
---
- # 未上市公司研究:多维度深度研究框架
+ # Private-Company Research: Multi-Lens Deep Framework
- 对一家未上市公司(如蚂蚁、小红书、SpaceX、Stripe)做深度研究。
+ Deep research on an unlisted company (e.g. Ant Group, ByteDance, SpaceX, Stripe).
- **最终目标**:在信息天然稀缺的条件下,尽可能还原这家公司的**真实价值**——不是市场给的估值,而是生意本身值多少钱。
+ **Ultimate goal**: under information scarcity, recover the company's **true value** — not the market valuation, but what the business is actually worth.
- ## 框架特点
+ ## Framework Characteristics
- 未上市 vs 上市研究的核心差异:无标准化财报(需多源拼凑、交叉验证);估值锚定少(依赖融资轮次、可比公司法、情景推演);信息不对称大(需"拼图式"研究);退出路径不确定(IPO/并购/二级转让均有可能)。
+ Private vs public research: no standardized financials (multi-source patchwork + cross-validation); few valuation anchors (funding rounds, comparables, scenarios); large information asymmetry ("jigsaw" research); uncertain exit path (IPO / M&A / secondary).
- ## AI 研究偏见自觉(核心前提)
+ ## AI Research Bias Self-Check (core premise)
- 未上市公司是 AI 研究偏见最严重的领域。警惕:
+ Private companies are where AI bias is worst. Watch for:
- - **虚假保守**——资料少时 AI 倾向给保守/模糊结论,但资料少 ≠ 公司不好。
- - **虚假精确**——为填满模板,AI 可能用"合理推测"伪装成"有据分析"。
- - **对标陷阱**——强行与上市公司对标时,继承上市估值逻辑,忽略独特价值。
- - **幸存者偏差**——网上能搜到的多是公司主动传播的好消息。
+ - **False conservatism** — with little data, AI gives conservative/vague conclusions, but scarce data ≠ bad company.
+ - **False precision** — to fill the template, AI disguises "reasonable guess" as "sourced analysis".
+ - **Comparables trap** — forcing a public-comp overlay inherits public-market logic and misses private-specific value.
+ - **Survivorship bias** — what's searchable online is mostly company-propagated good news.
- **应对**:宁可留白说"不知道",也不用推测填满表格伪装确定性;每个数据点标注置信度(🟢高/🟡中/🔴低);区分"可验证的事实"与"AI 的推理";信息极度稀缺时切换"第一性原理模式",只回答:① 这门生意解决什么真实问题?② 为什么是这个团队?③ 成功天花板 / 失败死点?④ 当前关键验证节点?
+ **Counter**: prefer leaving blanks ("I don't know") over filling tables with speculation to fake certainty; label every data point with confidence (🟢high/🟡medium/🔴low); separate verifiable fact from inference; when information is extremely scarce, switch to "first-principles mode" and answer only: ① what real problem does this business solve? ② why this team? ③ ceiling if it succeeds / how it dies if it fails? ④ the key validation node at this stage?
- > 反向利用信息不对称:市场对未上市公司信息少 → 定价效率低 → 恰恰可能是超额收益来源。
+ > Invert the asymmetry: the market knows little about private companies → pricing is inefficient → that's exactly where alpha may live.
- ## 执行编排
+ ## Execution
- 六个维度建议**并行**研究(用 `run_swarm`,每个 worker 一个维度;或顺序用 `web_search` / `web_search` 工具族逐维研究):
+ Six lenses, best run **in parallel** (via `run_swarm`, one worker per lens; or sequentially via `web_search`):
- | 角色 | 核心视角 |
- |------|----------|
- | business-decoder | 商业模式拆解 + 产品/用户分析:"这门生意的本质是什么" |
- | financial-detective | 财务数据拼凑 + 估值推演:"信息缺失下还原真实财务面貌" |
- | competitive-mapper | 行业格局 + 竞争 + 替代威胁:"谁在竞争,谁可能颠覆" |
- | risk-governance-analyst | 风险全景 + 管理层/治理/投资人:"什么可能出错,谁在掌舵" |
- | tech-ip-analyst | 技术栈/专利/研发/护城河:"技术壁垒真假与持久性" |
- | signal-miner | 替代数据:招聘/专利/诉讼/App/供应链:"常规信息外的蛛丝马迹" |
+ | Role | Lens |
+ |------|------|
+ | business-decoder | Business model + product/user analysis: "what is this business, essentially" |
+ | financial-detective | Financial patchwork + valuation: "recover the true financial picture under missing data" |
+ | competitive-mapper | Industry + competition + substitution: "who competes, who could disrupt" |
+ | risk-governance-analyst | Risk全景 + management/governance/investors: "what could go wrong, who's at the helm" |
+ | tech-ip-analyst | Tech stack / patents / R&D / moat: "is the tech barrier real and durable" |
+ | signal-miner | Alternative data (hiring / patents / litigation / app / supply chain): "clues beyond the usual sources" |
- team-lead(你自己)负责统筹、信息拼图、交叉验证、输出最终报告。
+ You (team-lead) integrate, patch the picture, cross-validate, output the final report.
- ## 维度 1:商业模式与用户(business-decoder)
+ ## Lens 1: Business Model & Users (business-decoder)
- - **核心生意定义**:一句话本质(段永平式:用朴素语言讲给聪明但不懂行的人)。解决什么问题?为谁创造价值?公司不存在用户会怎样?需求刚性(下行时会被砍吗)?
- - **收入模式**:广告/佣金/订阅/抽成/金融/SaaS/硬件/授权;各线占比与趋势;变现效率(ARPU/Take Rate/转化率);经常性 vs 一次性占比;集中度;收入确认是否激进。
- - **单位经济(U E)**:CAC(付费 vs 自然,分渠道,趋势)、LTV、LTV/CAC、回本周期、边际成本、规模经济拐点。
- - **产品矩阵与飞轮**:核心+延伸+孵化;网络/数据/规模飞轮如何运转;产品迭代速度。
- - **用户分析**:MAU/DAU(从 QuestMobile/Sensor Tower/SimilarWeb 推算)、S 曲线阶段、DAU/MAU、留存(次/7/30 日)、画像、口碑(App Store 评分趋势、社交舆情)。
- - **护城河(6 维 ★1-5)**:网络效应、转换成本、品牌心智、数据壁垒、监管牌照、规模经济——每项给证据 + 趋势(加宽/稳定/收窄)+ 持久性。综合评级:宽/中/窄/无。
+ - **Core business definition**: one sentence (Duan Yongping style: plain language to a smart layperson). What problem? For whom? If the company didn't exist, what would users do? Is demand rigid (cut in a downturn)?
+ - **Revenue model**: ads/commission/subscription/take-rate/financial/SaaS/hardware; mix and trend; monetization efficiency (ARPU / take rate / conversion); recurring vs one-off; concentration; predictability.
+ - **Unit economics**: CAC (paid vs organic, by channel, trend), LTV, LTV/CAC, payback, marginal cost, scale-inflection point.
+ - **Product matrix & flywheel**: core + extension + incubation; network/data/scale flywheel; iteration speed.
+ - **Users**: MAU/DAU (from QuestMobile/Sensor Tower/SimilarWeb), S-curve stage, stickiness (DAU/MAU, retention), profile, reputation (App Store trend, social sentiment).
+ - **Moat (6 dimensions, ★1-5)**: network effects, switching costs, brand mind, data barrier, regulatory license, scale economies — each with evidence + trend (widening/stable/narrowing) + durability. Overall: wide/narrow/none.
- ## 维度 2:财务数据拼凑与估值(financial-detective)
+ ## Lens 2: Financial Detective
- 未上市公司没有标准财报,多源拼凑、交叉验证,**每个数据标明来源、时间、置信度、推算方法**。
+ No standard financials; multi-source patchwork + cross-validation. **Every data point: source, time, confidence, derivation.**
- **数据来源优先级**:🟢 招股书/监管文件、母公司年报关联披露、监管处罚文件、债券/ABS 说明书 → 🟡 工商登记、融资新闻、第三方研报、深度媒体(晚点/The Information/36 氪/Bloomberg)→ 🔴 行业推算、前员工爆料。
+ **Source priority**: 🟢 prospectus/regulatory filings, parent-company annual report disclosure, regulatory penalties, bond/ABS offering documents → 🟡 business registry, funding news, third-party reports, deep media (LatePost/The Information/36Kr/Bloomberg) → 🔴 industry extrapolation, ex-employee leaks.
- **关键指标推算**:收入(规模/增速/结构/量×价分解)、成本(毛利/研发/销售/管理费用率,对标同业)、利润(EBITDA/净利/盈利时间线)、现金流(经营现金流/烧钱速度/现金跑道)、效率(人均收入/资本效率)。
+ **Key metrics**: revenue (scale/growth/mix/quantity×price), cost (gross margin/R&D/sales/G&A rates, vs peers), profit (EBITDA/net income/profitability timeline), cash flow (operating/burn rate/runway), efficiency (per-capita revenue, capital efficiency).
- **交叉验证**:同一指标多来源全部列出,不同方法推算检查是否收敛,标注"孤证"数据。
+ **Cross-validation**: list every source for the same metric; check convergence across methods; flag single-source ("isolated evidence") data.
- **融资历史**:完整时间线(轮次/金额/估值/领投),分析估值增长曲线健康度、融资间隔、是否 Down Round、老股东是否持续加注、最新轮条款推测(优先清算权/反稀释/对赌)对普通股价值的影响。
+ **Funding history**: full timeline (round/amount/valuation/lead investor); health of the curve, interval, down-rounds, whether existing investors keep participating; latest-round terms (liquidation preference / anti-dilution / ratchet) and their effect on common-share value.
- **估值(多方法交叉)**:① 最近融资估值法(考虑优先清算权打折 20-40%);② 可比上市公司法(选 3-5 家,对比 PS/PE/EV/EBITDA,加流动性折价 20-30%);③ DCF 情景(悲/中/乐,每假设有论据);④ 终局市值倒推(5/10 年终态 → 倒推当前合理区间 → 隐含年化回报);⑤ 交易对标法(近 2 年同业并购/融资倍数)。
+ **Valuation (multi-method)**: ① last-round (adjust for liquidation prefs, 20-40% discount); ② comparable public comps (3-5, PS/PE/EV-EBITDA, liquidity discount 20-30%); ③ DCF scenarios (bear/base/bull, each assumption grounded); ④ terminal-value rollback (5/10y terminal state → implied IRR); ⑤ transaction comps (recent M&A/funding multiples).
- **估值综合**:各方法是否收敛?差异大分析原因。区分"合理估值"和"保守估值(安全边际)"。
+ **Valuation synthesis**: do the methods converge? If divergent, explain. Distinguish "fair value" and "conservative (margin-of-safety) value".
- ## 维度 3:行业与竞争(competitive-mapper)
+ ## Lens 3: Competitive Landscape (competitive-mapper)
- - **市场**:TAM/SAM/SOM、渗透率、行业阶段(萌芽/成长/成熟/衰退)、增长驱动。
- - **产业链全景**(文本图):上游→公司环节(利润池占比)→下游;议价能力;结构性变化。
- - **波特五力**(★1-5):行业内竞争、新进入者、替代品、供应商、买方。
- - **竞争扫描**:直接/间接/替代/潜在进入者(巨头)——份额、融资、优势、劣势、威胁等级。2-3 个最直接对手多维对比。
- - **竞争动态**:近 12 月变化;从招聘/专利/产品推推断对手战略;技术(尤其 AI)与监管对格局的影响;赢家通吃 vs 多寡头。
- - **情景推演**:公司胜出 / 均势共存 / 被颠覆,各自条件与概率。
- - **全球对标**:海外对标公司的发展路径、估值、上市表现,及对标的局限性。
+ - **Market**: TAM/SAM/SOM, penetration, stage (emergence/growth/mature/decline), growth drivers.
+ - **Value chain** (text map): upstream → company's link (profit pool share) → downstream; bargaining power; structural shifts.
+ - **Porter's five forces** (★1-5): rivalry, new entrants, substitutes, supplier power, buyer power.
+ - **Competitor scan**: direct/indirect/substitute/potential entrants (giants) — share, funding, strengths, weaknesses, threat level. Multi-dimensional compare with 2-3 closest competitors.
+ - **Dynamics**: last-12-month changes; infer competitor strategy from hiring/patents/products; tech (esp. AI) and regulation effects; winner-take-all vs oligopoly.
+ - **Scenarios**: company wins / coexistence / disrupted — conditions and probabilities.
+ - **Global benchmarks**: overseas analogs, path, valuation, post-IPO performance, and benchmark limitations.
- ## 维度 4:风险与治理(risk-governance-analyst)
+ ## Lens 4: Risk & Governance (risk-governance-analyst)
- - **创始人/CEO**:背景、战略眼光(过去 3 年预判准确度)、执行力(承诺兑现)、格局价值观、争议事件。★1-5。
- - **核心团队**:高管背景、近 2 年人才流动(谁走/谁来/净流入)、互补性、文化信号(Glassdoor/脉脉,趋势比绝对值重要)、关键人依赖度。
- - **股权与治理**:创始人控制权(AB 股/一致行动/VIE)、稀释趋势、员工持股;董事会构成、关联交易、同业竞争、大小股东利益冲突。
- - **投资人阵容**:领投品牌背书、产业资本协同、退出压力(基金存续期/二级抛售/对赌到期)、红旗信号(声誉不佳投资人/早期清仓/老股东不跟投)。
- - **全维度风险清单**:监管/竞争/技术/人才/融资/IPO/地缘/商业化/治理/合规/宏观/ESG——每项概率×影响×严重度×可对冲性×监控指标。
- - **退出路径**:A/港/美 IPO、并购、二级转让、SPAC、长期不退出——可能性、时间窗、估值、前提、障碍。
- - **最坏情景(芒格逆向)**:3 种具体失败路径及概率;清算价值;聪明人为什么不投(至少 5 条);类似公司失败案例;"论文已破"的信号。
+ - **Founder/CEO**: background, foresight (3-year prediction accuracy), execution (promise delivery), values, controversies. ★1-5.
+ - **Core team**: backgrounds, 2-year talent flow (who left/joined, net), complementarity, culture signals (Glassdoor trends), key-person dependency.
+ - **Equity & governance**: founder control (dual-class / concerted action / VIE), dilution trend, employee equity; board, related-party, same-industry competition, majority/minority conflicts.
+ - **Investor roster**: lead-investor brand, strategic capital synergy, exit pressure (fund life / secondary sales / ratchet maturity), red flags.
+ - **Risk matrix**: regulatory / competition / tech / talent / funding / IPO / geopolitical / monetization / governance / compliance / macro / ESG — each probability × impact × severity × hedgeability × monitor.
+ - **Exit paths**: A/HK/US IPO, M&A, secondary, SPAC, stay-private — probability, window, valuation, preconditions, obstacles.
+ - **Worst case (Munger inversion)**: 3 specific failure paths + probabilities; liquidation value; why smart money doesn't invest (≥5 reasons); failed analogs; the "thesis broken" signal.
- ## 维度 5:技术与知识产权(tech-ip-analyst)
+ ## Lens 5: Tech & IP (tech-ip-analyst)
- - **技术栈**:从招聘/博客/开源/演讲推断;技术债信号(大量招重构岗、栈切换、故障投诉)。
- - **专利**(Google Patents/CNIPA/USPTO):总数/申请中/近 2 年新增/领域分布/引用/国际布局;质量(核心专利?领域匹配?诉讼?);趋势;对手对比。
- - **研发**:投入(人数/费用率,对标)、产出(论文/会议/开源/博客)、效率(研发到落地速度、商业转化率)。
- - **技术人才**:核心领导者背景、人才密度(顶尖机构占比)、薪酬竞争力、流失信号、招聘方向(反映战略)。
- - **技术护城河(★1-5)**:算法/模型、数据、工程、人才、生态——每项持久性。AI 时代技术壁垒半衰期可能很短。
- - **AI/新技术影响**:公司是 AI 变革受益者还是被颠覆者?
+ - **Tech stack**: inferred from hiring/blog/open-source/talks; tech-debt signals (refactor hiring, stack switches, outage complaints).
+ - **Patents** (Google Patents/CNIPA/USPTO): total/pending/last-2y/field/citation/international; quality (core patents? aligned to business? litigation?); trend; vs competitors.
+ - **R&D**: investment (headcount/expense rate vs peers), output (papers/conferences/open-source/blog), efficiency (research-to-product, commercialization).
+ - **Tech talent**: core leaders' backgrounds, density (top-institution share), comp competitiveness, attrition signals, hiring direction (→ strategy).
+ - **Tech moat (★1-5)**: algorithm/model, data, engineering, talent, ecosystem — each with durability (AI-era half-life may be short).
+ - **AI/new-tech impact**: is the company a beneficiary or a target of disruption?
- ## 维度 6:替代数据信号(signal-miner)
+ ## Lens 6: Alternative-Data Signals (signal-miner)
- > 未上市公司常规信息有限,替代数据往往比新闻报道更接近真实经营。
+ > Private companies have limited conventional info; alt-data often beats news.
- - **招聘**(LinkedIn/Boss/脉脉/Indeed):规模趋势、结构(研发/产品/销售/数据/国际化/合规/IR)——招 IR = IPO 信号;招合规 = 监管或 IPO;岗位 JD 技术栈 = 战略方向。
- - **App/产品**(App Store/七麦/SimilarWeb):排名、评分趋势、下载量、更新频率、差评集中点、Web 流量。
- - **舆情**(微博/知乎/小红书/X/Reddit):官方互动数据、自发讨论热度情感、KOL 评价、负面事件、内部人爆料。
- - **工商法律**(天眼查/企查查):注册资本/实缴/股东变更/子公司(新设=新业务,注销=收缩)/经营范围变更;诉讼/仲裁/处罚/被执行记录。
- - **供应链**:核心供应商(上市则查财报披露)、招标采购、合作评价。
- - **数字足迹**:注册域名(新域=新业务)、子域名(api/pay 暗示架构)、商标(新品牌)。
- - **行业曝光**:高管会议演讲、奖项榜单、政府/协会互动、媒体曝光频率质量。
- - **二级交易信号**(如有):SharesPost/EquityZen/微信群老股,隐含估值 vs 最近融资,员工抛售期权。
+ - **Hiring** (LinkedIn/Boss/Indeed): scale/trend, structure (R&D/product/sales/data/international/compliance/IR — IR hiring = IPO signal; compliance = regulatory or IPO; JD tech stack = strategy).
+ - **App/product** (App Store/七麦/SimilarWeb): rank, rating trend, downloads, update frequency, complaint themes, web traffic.
+ - **Social sentiment** (Weibo/Zhihu/Xiaohongshu/X/Reddit): official engagement, organic discussion, KOL views, negative events, insider leaks.
+ - **Business/legal** (天眼查/企查查): registry/paid-in/equity changes/subsidiaries (new = new biz; deregistered = contraction)/scope changes; litigation/arbitration/penalties/enforcement.
+ - **Supply chain**: known suppliers (if listed, check their filings), procurement, partner evaluation.
+ - **Digital footprint**: registered domains (new = new biz), subdomains (api/pay → architecture), trademarks (new brands).
+ - **Industry exposure**: exec talks, awards, government/association interaction, media frequency/quality.
+ - **Secondary-market signals** (if any): SharesPost/EquityZen, implied valuation vs last round, employee selling.
- **反常信号清单(最重要)**:与对外叙事不一致、与行业常识不符、突然变化(招聘急缩/高管密集离职)、无法解释的现象。
+ **Anomaly list (most important)**: things inconsistent with the company's narrative; inconsistent with industry norms; sudden changes (hiring freeze / executive departures); unexplained.
- ## 交叉验证(team-lead 必做)
+ ## Cross-Validation (team-lead, mandatory)
- 汇总前,team-lead 必须:
+ Before synthesis, the team-lead must:
- 1. **数据冲突仲裁**:同一指标不同来源是否一致?冲突时列出全部,说明采信哪个及理由。
- 2. **信号一致性检验**(矩阵):业务增长 vs 招聘趋势?技术领先叙事 vs 专利/人才数据?估值水平 vs 竞争地位?管理层叙事 vs 实际行动信号?(矛盾要解释)
- 3. **信息拼图**:白区(已知)/灰区(有线索不确定)/黑区(完全未知)。
- 4. **反偏见检查**:是否"正面详、负面略"?每个正面判断是否有反面检验?
+ 1. **Data conflict arbitration**: same metric across sources — list all, state which is adopted and why.
+ 2. **Signal consistency matrix**: business-growth signal vs hiring trend? tech-leadership narrative vs patent/talent data? valuation level vs competitive position? management narrative vs action signals? (contradictions must be explained)
+ 3. **Information jigsaw**: white zones (known) / gray (clues but uncertain) / black (unknown).
+ 4. **Bias check**: is positive info detailed while negative is brief? Does every positive judgment have a reverse check?
- ## 最终报告结构
+ ## Final Report Structure
- 1. **一句话结论**(50-100 字):生意值多少钱,为什么。
- 2. **公司画像速览**(含置信度列)。
- 3. **六维评分总表**(每维 ★1-5 + 核心判断 + 置信度 + 信息完整度),综合评分。
- 4. **关键数据拼图**(仅交叉验证后的,含来源数量 + 置信度)。
- 5. **信号一致性矩阵**。
- 6. **各维度分析摘要**(每维 3-5 条最重要发现)。
- 7. **真实价值评估**:生意本质 + 护城河评分卡(7 维)+ 估值(5 方法)+ **综合真实价值区间**(保守/合理/乐观 + 当前市场估值 + 安全边际%)。
- 8. **投资论点**:看多 5-7 条 vs 看空 5-7 条(每条附来源),哪方更有说服力。
- 9. **风险矩阵**(Top 3 + 应对)。
- 10. **退出路径评估**。
- 11. **投资决策表**(一页纸:核心逻辑 3 句 + 价值区间 + 关键假设与验证节点 + 致命风险与"论文已破"信号 + 结论 + 预期回报/时间)。
- 12. **信息盲区地图**(维度 / 已知 / 缺失 / 缺失影响 / 获取建议):缺失是否影响核心结论?影响则明确"在 X 缺失下,结论置信度为 Y"。
- 13. **持续跟踪清单**(事项/频率/来源/指标/预警阈值)。
- 14. **总结段落**(150-250 字)。
+ 1. **One-line conclusion** (50-100 words): what's it worth, why.
+ 2. **Company snapshot** (with confidence column).
+ 3. **Six-lens scorecard** (★1-5 + core judgment + confidence + completeness), overall score.
+ 4. **Key data jigsaw** (only cross-validated, with source count + confidence).
+ 5. **Signal-consistency matrix**.
+ 6. **Per-lens summary** (3-5 top findings each).
+ 7. **Fair value assessment**: business essence + 7-dimension moat card + 5-method valuation + **fair value range** (conservative/reasonable/optimistic + current market valuation + margin of safety %).
+ 8. **Investment thesis**: bull 5-7 (with sources) vs bear 5-7 (with sources), which side is stronger.
+ 9. **Risk matrix** (top 3 + mitigation).
+ 10. **Exit-path assessment**.
+ 11. **Investment decision table** (one-pager: core logic 3 sentences + value range + key assumptions & validation nodes + fatal risks & "thesis broken" signals + conclusion + expected return/timeline).
+ 12. **Information-gap map** (dimension / known / missing / missing-impact / how-to-get): does the gap affect the core conclusion? If yes, state "under missing X, conclusion confidence is Y".
+ 13. **Tracking checklist** (item / frequency / source / metric / alert threshold).
+ 14. **Summary paragraph** (150-250 words).
- 把报告用 `write_file` 写入 `reports/{公司}/{公司}-private-{YYYYMMDD}.md`。可用 `report_audit` 对其中的数字做准出门。
+ Save via `write_file` to `reports/{company}/{company}-private-{YYYYMMDD}.md`. Run `report_audit` on the numbers as a quality gate.
- ## 数据标注规范(严格执行)
+ ## Data Labeling Standard (strict)
- - 每个关键数据标注**来源**(具体到媒体 + 文章)、**时间**(年月)、**置信度**(🟢招股书/官方 / 🟡可信媒体研报 / 🔴推算传闻)。
- - 冲突数据**全部列出**并说明差异和采信判断。
- - **区分事实与推理**:事实正常字体;推理/推算用*斜体*并标注推算方法。
- - 无法获取的信息明确标"数据缺失",不编造。
+ - Every key data point: **source** (specific to media + article), **time** (year/month), **confidence** (🟢 prospectus/official / 🟡 credible media / 🔴 estimate/rumor).
+ - Conflicting data: **list all** + explain difference and adoption.
+ - **Separate fact from inference**: fact in normal text; inference in *italics* with derivation.
+ - Missing info: explicitly mark "data missing"; never fabricate.
- ## 重要原则
+ ## Key Principles
- 1. **6 维度并行**(用 `run_swarm` 同一条消息并行,或顺序)。
- 2. **推算要透明**——展示计算逻辑和每步假设,不凭空给数字。
- 3. **交叉验证**——关键数据至少 2 来源,冲突都列出。
- 4. **信号一致性检验**——汇总阶段必做跨维度一致性检查。
- 5. **结论要明确**——不回避投资/观望/回避建议,但说明置信度。
- 6. **中英文搜索**——未上市公司信息分布中英文媒体。
- 7. **诚实留白**——区分"有据分析"与"推测填充",允许"此维度数据不足,无法给出有意义结论"。
- 8. **替代数据不是噪音**——招聘/专利/诉讼/App 可能比新闻更接近真实经营。
- 9. **真实价值导向**——目标是判断生意值多少钱,不是输出好看报告。信息不足以估值时,直接说"信息不足,无法给出可靠估值"。
- 10. **资料少 ≠ 公司不好**——AI 分析篇幅短 ≠ 投资确定性低;信息极度稀缺时切第一性原理模式。
+ 1. **6 lenses in parallel** (`run_swarm`, or sequential).
+ 2. **Transparent derivation** — show the math and assumptions; don't hand-wave numbers.
+ 3. **Cross-validate** — key data ≥2 sources; conflicts all listed.
+ 4. **Signal-consistency check** — mandatory cross-lens check at synthesis.
+ 5. **Clear conclusion** — don't dodge invest/watch/avoid; state confidence.
+ 6. **Search in both EN and CN** — private-company info spans both.
+ 7. **Honest blanks** — distinguish "sourced analysis" from "speculative fill"; "this dimension lacks data, no meaningful conclusion" is acceptable.
+ 8. **Alt-data is not noise** — hiring/patents/litigation/app data may be closer to truth than news.
+ 9. **True-value focus** — the goal is what the business is worth, not a pretty report. If info can't support a reliable valuation, say so.
+ 10. **Scarce data ≠ bad company** — short AI output ≠ low certainty. Under extreme scarcity, switch to first-principles mode.