Kathy Xu

1. Snapshot and testable edge thesis

Mechanism, predictions, and falsifiers

The mechanism is category map → consumer truth → founder learning and integrity → unit-economics and leadership path → sufficiently large initial check → selective follow-ons → operating help → patient but opportunity-cost-aware hold. [researcher inference | moderate] The sequence is a synthesis of Xu's public doctrine, not a recovered Capital Today investment memo. 2018 long-form interview, method and case discussions 2023 public remarks, value, venture, and concentration

If the edge is real, unseen files should contain: category censuses before company selection; verbatim consumer evidence that changes a forecast; references testing founder integrity and learning rather than charisma; explicit customer-acquisition payback and cash-runway gates; larger ownership or reserves when conviction rises; and written reasons to stop, sell, or decline. It weakens if famous outcomes were mostly privileged access, founder execution, team decisions, cheap-growth beta, or post-outcome narrative. It fails as an economic claim if complete, vintage-adjusted net cash flows do not outperform after the largest winner is removed. [researcher inference | strong]

2. Core philosophy

Xu's public system has six linked propositions:

1. Study the category before admiring the company. [investor-stated | strong] She repeatedly distinguishes the “forest” from the attractive “tree” and describes studying prior category leaders. 2015 speech transcript, errors and category mapping
2. Get close to the consumer. [investor-stated | strong] Published accounts describe direct interviews and household observation, including a reported 2,000 interviews after missing Pinduoduo's early thesis. The count is self-reported and not independently audited. 2022 Nanjing University alumni interview
3. Back a founder who learns, keeps promises, and attracts talent. [investor-stated | strong] Xu treats integrity as a threshold and learning velocity as a compounding asset. 2018 long-form interview
4. Choose markets where frequency, necessity, and scale can create a category leader. [investor-stated | moderate] This is clearest in her retail, marketplace, and local-services examples; it is a heuristic, not a demonstrated universal law. Chaos Academy transcript
5. Concentrate and finance the strongest cases through difficult periods. [investor-stated | moderate] Xu has described a six-to-eight-company attention capacity and top positions dominating the portfolio. Public filings show concentration in a narrow U.S.-listed slice, not the complete fund. 2025 methodology account Capital Today Q2 2026 13F cover
6. Treat price, competence, and downside more explicitly than the early rhetoric did. [investor-stated | strong] Her 2025 essay emphasizes competence circles, discounted cash flow, and margin of safety. Xu-authored 2025 essay, value-investing headings

Synthesis. [researcher inference | strong] Xu is best read neither as a pure growth investor nor as a Buffett clone. Her operative style is concentrated private-company underwriting in consumer and internet categories, later expressed through value-investing language. The durable core is category leadership plus founder learning; the important revision is that growth must sit inside a competence, cash-flow, and price discipline.

3. Sourcing

For each opportunity, a usable sourcing log is: identity and aliases; originator; relationship duration; category census; why now; consumer evidence; conflict; competing process; allocation; pass/re-entry rule. The identity field matters unusually here: Xu Xin / 徐新 must not be confused with Xu Xiaoping (Bob Xu) or JD.com executive Sandy Ran Xu. [researcher inference | strong] KANZHUN Form 20-F

4. Picking

Decision sequence

1. Define the market in customer terms. Measure frequency, necessity, addressable spend, fragmentation, and the reason a leader can emerge. [researcher inference | moderate] Chaos Academy transcript
2. Map the forest. List incumbents, startups, substitutes, historical analogues, and failure modes before meeting the favored founder. [investor-stated | strong] 2015 speech transcript
3. Triangulate. Reconcile management claims with customers, former employees, competitors, suppliers, and observable behavior. [researcher inference | strong]
4. Test founder quality as behavior. Ask for examples of truth-telling under pressure, learning from error, recruiting stronger executives, and allocating scarce cash. [researcher inference | moderate] 2018 interview
5. Underwrite three controls. Xu's recurring frame is supply/product control, channel access, and customer mindshare/brand. [investor-stated | moderate] Chaos Academy transcript
6. Translate growth into cash. Model acquisition payback, contribution margin, working capital, fixed costs, and the funding path. Xu publicly used a two-year customer-acquisition payback preference, with three years as an outer bound in one BOSS-related account; it is not shown as a universal firm rule. [investor-stated | moderate] 36Kr BOSS account
7. Price uncertainty. State the competence gap, downside value, dilution, and return at base, bear, and bull outcomes. [researcher inference | strong] Xu-authored 2025 essay

Red flags and stop conditions

5. Deal and portfolio mechanics

The record is strongest on legal holdings and governance at specific dates, weaker on sale timing, and weakest on complete investor cash flows. For JD.com, for example, the prospectus and 13G establish ownership snapshots and control chains; they do not verify the widely repeated 150–160x narrative. [researcher inference | strong] JD.com Schedule 13G

6. Vehicle and incentive fit

Xu's philosophy cannot be separated from the vehicles that make it feasible. Capital Today lists Fund I (2006), Fund II (2010), and evergreen tranches from 2014 onward, with officially reported evergreen tranche fund sizes totaling $2.849 billion. This sum is not current AUM or NAV and is not an audited net-performance record. Capital Today

Operating rule. [researcher inference | strong] Before transferring a Xu-style practice, write a one-page vehicle-fit memo covering fund term, ownership target, reserves, attention capacity, follow-on decision rights, cross-fund allocation, valuation, currency, regulatory, and liquidity constraints. If the vehicle cannot survive the proposed hold without forced financing or conflicted transfer, the strategy does not fit.

7. Team attribution and founder relationships

The public narrative is unusually person-centered, but the assets were produced by teams. Capital Today professionals, earlier Baring partners, company founders, co-investors, directors, executives, lenders, and public shareholders all contributed. KANZHUN's filing proves Xu's directorship and career biography; JD.com's filings show a GP control chain; neither turns all company value creation into personal attribution. KANZHUN Form 20-F JD.com IPO prospectus

For every case, allocate six functions: origination; diligence/IC; security and allocation; board/governance; recruiting/operating help; sale/distribution. Use personal, firm, earlier employer, and unclear as separate ownership buckets. NetEase, Vanda, Comtech, and other pre-2005 roles must not be silently counted as Capital Today deals; ChinaHR's initial personal investment must not become a fund investment without documentation. [researcher inference | strong]

The founder rule is behavioral: integrity, capacity to learn, ability to recruit, and persistence. It is not a license for founder exceptionalism. Zhen Kungfu demonstrates why personal confidence must be paired with controls; the later criminal judgment concerned the founder and does not establish investor misconduct. government case record Guangzhou court, shareholder context

8. Views on the craft

Xu treats investing as a small-team craft built from close observation, repeated pattern study, concentrated attention, and self-correction. Her reported practice of recording investment discussions and holding multi-day annual postmortems is among the most transferable parts of the record because it creates a contemporaneous decision trail. [investor-stated | moderate] 2015 speech transcript

The craft has four disciplines:

• Epistemic: distinguish what the investor saw before investment from what success later made salient.
• Economic: connect market love and founder admiration to ownership-adjusted cash returns.
• Organizational: reserve partner attention for the small number of cases where intervention has a mechanism.
• Moral and legal: test integrity with controls; preserve allegation, finding, and exoneration boundaries.

The public corpus is rich in polished retrospection and thin in losing memos, rejected references, fund audits, and dissent. An analyst should therefore learn more from the questions and record-keeping system than from the celebratory return numbers. [researcher inference | strong]

9. Signature bets and repeatable process

These cases support process reconstruction, not a league table. The repeatable artifact is a pre-decision lab: make the reader decide invest/pass; check and ownership; milestones; kill criteria; reserves; governance; exit evidence before revealing the outcome. [researcher inference | strong]

10. Mental models and maxims

The distinct models are developed in the mental-model field guide:

1. Map the forest before choosing the tree.
2. Earn a competence circle through consumer immersion.
3. Underwrite founder learning and integrity as behaviors.
4. Capitalize a plausible category leader—but only through evidence gates.
5. Seek control of product, channel, and customer mindshare.
6. Concentrate for home runs within a hard attention budget.
7. Finance through winters with prewritten milestones.
8. Treat holding and liquidity as a second investment decision.
9. Use recruiting, metrics, and brand as bounded investor interventions.
10. Convert mistakes into a recorded decision-and-postmortem loop.
11. Require price protection inside the competence circle.

The maxims are useful only in conditional form. “Category leader” without unit economics invites subsidy races; “great founder” without controls invites governance failure; “price does not matter” fails outside rare outliers; “patience” without stop conditions becomes escalation; and “consumer truth” without sampling discipline becomes anecdote. [researcher inference | strong]

11. Evolution and contradictions

Three contradictions remain unresolved: high concentration versus incomplete denominator; hands-on help versus unclear causal attribution; and long holding versus the need to return capital and reassess price. Those tensions should stay visible rather than being reconciled by rhetoric. [researcher inference | strong]

12. Failures, misses, and limits

• ByteDance pass. [investor-stated | moderate] Xu says she did not understand feed advertising at the relevant stage. This is a competence-boundary miss, not evidence that she could have obtained an allocation. 2025 methodology account
• Pinduoduo pass. [investor-stated | strong] She attributes the early miss to distance from lower-tier consumers. Later interviews and a public holding show changed behavior, not recovery of the forgone private return. 2022 alumni interview
• Yonghe early exit. [retrospective witness | moderate] The lesson is that selling a winner can be as consequential as choosing it; complete cash flows are unavailable. 2015 speech
• ChinaHR. Xu is reported to have invested household savings, and a 2007 profile reports roughly 800x paper value. [retrospective witness | provisional] No personal cost, shares-sold, cash-received, dilution, or vehicle-chain ledger supports a cash return. The later company deterioration is a separate company outcome. [researcher inference | strong] 2007 Sina profile 2014 commencement address
• Yonghui Yunchuang. [documented behavior | moderate] Later losses and store contraction challenge a simple “new retail” growth story and illustrate how a category thesis can outpace unit economics. The Paper, store closures and losses
• Zhen Kungfu. [contemporaneous record | strong] Founder criminal conduct and shareholder conflict test founder diligence and board control. They do not establish wrongdoing by Xu or a known fund loss without transaction records. government case record Guangzhou court, shareholder context
• Inoherb dispute. [contemporaneous record | moderate] Press allegations about founder/shareholder conflict require allegation labels; no cited final adjudication supports a stronger claim. Yicai report

The larger limitation is selection bias. Public narratives foreground JD.com, Meituan, and a few misses but do not disclose every investment, pass, write-off, continuation, or distribution. The philosophy is therefore teachable as a hypothesis-generating system, not verified as a complete performance engine. [researcher inference | strong]

13. Historical operating environment

Skill claims must be regime-adjusted. Scarce capital, China consumer growth, mobile adoption, permissive subsidy competition, public-market windows, and regulatory pivots can make the same behavior look brilliant in one vintage and reckless in another. [researcher inference | strong]

14. Modern translation as of 2026-08-23

The translation preserves Xu's learning loop while removing China-2000s assumptions. It is not advice to copy her sector, geography, check size, or public/private crossover. [researcher inference | strong]

15. Comparison with completed peers

Only completed canon profiles are used here.

16. The strongest case against greatness

The skeptical case is that Xu is an unusually effective storyteller and company counselor whose visible record is dominated by a favorable China consumer-internet regime, privileged access, founder execution, team capital, and a few memorable outliers. JD.com return numbers are repeated without a complete cash-flow ledger. Meituan was a post-merger underwriting rather than proof that Xu originated the eventual winner at formation. NetEase belonged to Baring. ChinaHR began as a personal investment. The current 13F covers only a small, reportable public slice. [researcher inference | strong]

The doctrine may also be more coherent after the fact than in real time. “Price is secondary” and later margin-of-safety language do not automatically reconcile. Consumer immersion followed the Pinduoduo miss. Founder admiration coexists with Zhen Kungfu's governance failure. Aggressive follow-on support can be skill in JD.com and escalation in an ordinary company; without the denominator, the reader cannot tell how often each occurred. [researcher inference | strong]

The strongest answer would be discriminating evidence: complete fund cash flows; security-level cost, dilution, sale, and distribution ledgers; all investments, passes, and write-offs; original IC recordings and memos; named partner roles; board records; and matched results by stage, vintage, sector, ownership, and regulation. If those show that the stated method preceded decisions and worked beyond the top outlier, the edge strengthens. If they show ordinary net returns, team-owned decisions, and rules concentrated in winner stories, the skeptical thesis wins.

17. Study guide, glossary, and go deeper

Recommended reading path

1. Start with the profile for identity, chronology, people, and the falsifiable edge.
2. Read writings to separate Xu-authored doctrine, quoted public statements, and metadata-only material.
3. Read talks for direct voice, access grades, and contradictions across time.
4. Read investments before accepting any return story; focus on the denominator, role map, security evidence, and adverse cases.
5. Read adverse and miss postmortems before famous wins, then apply the mental-model field guide to a new company.
6. Use sources for locators, source-family independence, identity traps, and unresolved routes.

Exercises

1. Rebuild JD.com with only SEC filings. Mark each repeated return number verified, derived, claimed, or not computable.
2. Write a no-hindsight Pinduoduo pass memo, then a separate later public-purchase memo. Do not let later ownership repair the original decision.
3. Compare Meituan before and after the Dianping merger. Decide which opportunity Xu actually underwrote and what evidence would allocate credit.
4. Construct a ten-company category census and pre-register the three facts that would change the ranking.
5. Size a concentrated position, then rerun the fund with the top outcome at zero and follow-ons doubled.

Compact glossary

• Category census: a dated list of serious companies, substitutes, and historical analogues before selection.
• Competence circle: the domain where the team can identify what matters and what it does not know.
• Consumer immersion: structured direct observation used to test, not decorate, a market thesis.
• Three controls: product/supply, channel/distribution, and customer mindshare/brand.
• Attention budget: the maximum number of companies a partner can support without degrading decisions.
• Financing winter: a period when external capital is scarce and runway exposes business quality.
• Second decision: the fresh underwriting required to hold, sell, follow on, or transfer a position.
• Return bridge: dated cash flows from cost through sales/distributions, rather than headline valuation.
• Source-family independence: two sources with genuinely separate reporting or records, not copies of one interview.
• Identity fence: the names, entities, roles, and dates that prevent conflating people or institutions.

Sourcing checklist

• Identity and aliases verified in a primary filing.
• Originator, relationship, conflict, and allocation recorded.
• Full category census and pass/re-entry reasons preserved.
• Consumer sample and disconfirming evidence logged.

Investment checklist

• Frequency, necessity, scale, and category-leadership mechanism quantified.
• Founder integrity, learning, and recruiting tested behaviorally.
• Product, channel, and mindshare controls scored.
• Security, ownership, dilution, reserves, payback, runway, and regulatory risks modeled.

Portfolio and governance checklist

• Position fits loss budget, correlation, reserves, and partner attention.
• Every follow-on has a milestone, cumulative-exposure view, and stop rule.
• Board rights, related parties, data integrity, and escalation routes documented.
• Hold/sell decision is re-underwritten at each liquidity event.

Evidence checklist

• Every material claim has a claim-local full-page URL and locator where available.
• Investor statement and researcher inference use the exact two-axis labels.
• Round size, investor check, stake, value, realized proceeds, and net return remain distinct.
• Allegation, finding, acquittal/denial, and unresolved status remain distinct.

Ranked sources to go deeper

1. KANZHUN Form 20-F — primary biography and directorship anchor.
2. JD.com IPO prospectus — ownership, control, and risk evidence.
3. Meituan IPO prospectus — stake, corporate history, and listing context.
4. Capital Today — current firm description and vehicle claims.
5. Xu-authored 2025 essay — clearest late intellectual statement.
6. 2022 Nanjing University alumni interview — mistakes, consumer learning, and portfolio claims.
7. 2018 long-form interview — category, founders, company cases, and failure claims.
8. 2015 Chaos Academy transcript — consumer/channel doctrine and operating interventions.
9. 2015 speech transcript — misses, financing, founder time, and postmortem practice.
10. Capital Today Q2 2026 13F table — narrow public-portfolio snapshot with strict scope limits.

Pareto 80/20 — what drove the record

The public record does not support a literal 80/20 split, a career hit rate, or a ranking by net return. Capital Today does not publish the complete deal, ownership, disposition, fee, carry, and distribution ledger needed to compute MOIC, IRR, DPI, TVPI, or Kathy Xu's personal wealth. The three rankings below therefore answer different questions. They rank verification priority, historical importance, and learning value—not comparable cash returns. [researcher inference | strong]

Economic contribution — directional verification priority

Dominant-outlier test. JD is the strongest candidate for economic dominance, but the test cannot be quantified because the required fund cash-flow denominator is missing. Removing JD leaves Meituan, Yifeng, Three Squirrels, BOSS, and Xingsheng as plausible material contributors; their evidence is still not cash-flow comparable. The later claim that the top three to five deals generated about 90% of returns is Xu's retrospective generalization, not a published audited attribution analysis. [investor-stated | provisional] 36Kr interview, “three to five” concentration discussion

Historical importance — franchise and strategy inflections

Decision-learning value — where an investment committee should spend time

Evidence boundary and searched denominator

Scope, entities, years, aliases, and search method

This is an institution-association study spanning Xu's reported Peregrine and Baring work (approximately 1995–2005) and Capital Today from its 2005/2006 formation through the 2026-08-23 research cutoff. It separates Xu, the investment team, limited partners, Fund I, Fund II, the Evergreen tranches, continuation vehicles, nominee companies such as Best Alliance, co-investors, boards, founders, and portfolio companies. A corporate financing is not a Capital Today check; a company valuation is not fund proceeds; control of a GP is not personal ownership of portfolio assets. [researcher inference | strong]

Searches covered SEC, HKEX, SZSE, SSE/CNINFO and issuer prospectuses; shareholder agreements; ownership filings; Chinese regulator, court, and prosecutor records; company announcements; Capital Today's current and archived site; interviews in Chinese and English; and searches of English and Chinese aliases. Name variants included Kathy Xu, Hsin Kathy Xu, Xu Xin, 徐新, 今日资本, Capital Today, CTG, Capital Today China Growth, CTCG, CTG Evergreen, Best Alliance, and each company/legacy name below. [documented behavior | strong]

Inclusion rule. Count a company/opportunity once when a dated public route establishes a scoped vehicle association, a personal/Baring/Peregrine claimed deal path, or a specifically admitted miss that is essential to evaluate selection. Collapse renames and mergers into one path when they represent continuity; keep Dianping and the combined Meituan-Dianping path separate because they represent a pre-merger exposure and a later disclosed vehicle position. Include failures, misses, ordinary outcomes, continuing holdings, and disputed paths. Exclude a name when only a list-copy or general association remains and no deal-level route survives scrutiny. [researcher inference | strong]

Denominator reconciliation

The observed-source reconstruction ratio is 20 / 44 = 45.5%, but this is not a portfolio-coverage statistic. The full population of every Capital Today investment, write-off, follow-on, transfer, and disposition is not public; public lists and interviews select for survivors and stories. Career/fund coverage fraction: not computable. [researcher inference | strong]

Selection and evidence hazards

• Public filings overrepresent companies that reached an IPO, major transaction, or regulatory event. Private failures and small outcomes are under-observed. [researcher inference | strong]
• Xu's interviews are valuable for decision logic but are retrospective, promotional, and frequently omit the legal holder, syndicate allocation, dilution, and cash distributions. [researcher inference | strong]
• “Capital Today invested” does not establish that Xu sourced, approved, monitored, or personally profited from the deal. Team roles are attributed only where a filing or direct account supplies them. [researcher inference | strong]
• Numbers described as “return” in secondary profiles may instead be company value, gross security value, or a simple headline division. Every such number is quarantined below. [researcher inference | strong]

Historical market, regulatory, and base-rate context

The base-rate lesson is not that concentration automatically works. Consumer and internet successes are the easiest public stories to retrieve; failed private businesses, bridge rounds, early sales, and unrealized holdings are less visible. The correct prior is therefore unknown portfolio-wide loss frequency with severe selection bias, not a marketing-derived hit rate. [researcher inference | strong]

Vehicle economics and incentives

Vehicle chronology and what is actually known

Mergermarket reports—not an executed LPA reviewed here—that the Evergreen design spans 28 years in seven four-year periods and lets an LP decide at each interval whether to renew capital, seek liquidity, or let exposure run off. That account explains the intended duration/liquidity trade-off but remains source-reported and conditional. Public filings show offshore fund, GP, nominee, and holding-company layers; they do not disclose complete legal-form terms, partner geographic authority, country allocations, or the LP capital base. [retrospective witness | provisional] Mergermarket, structure discussion

The official site's displayed $280m + $408m + $752m + $989m + $1.108bn amounts must not be added and called current assets under management: vintages may overlap through investments, realizations, transfers, and continuation structures. Likewise, a $500m reported target is not capital raised. [researcher inference | strong]

Incentive questions that remain open

Long duration can align the vehicle with slow compounding, but only if LP liquidity, valuation, fee bases, carry crystallization, continuation pricing, conflicts, and succession are well governed. The public record does not disclose the compensation/carry split, management-company ownership over time, key-person clause, investment-committee voting, clawback, reserve formula, or succession plan. The website names a platform; filings reveal Xu, Wen Baoma, and Fang Min at particular GP layers; neither supports the claim that every investment or gain is Xu's alone. [researcher inference | strong]

The Q2 2026 Form 13F is a narrow current control: Capital Today Evergreen Fund reported three U.S.-listed long positions with aggregate fair value $129,498,857—PDD 1,385,004 ADS / $105,648,105, Dingdong 11,879,332 ADS / $22,927,111, and KANZHUN 71,767 ADS / $923,641. PDD is $105,648,105 / $129,498,857 = 81.6% of that filing's reported value. This is a researcher calculation from one quarter's 13F, not whole-fund concentration, cost, return, or Xu's personal portfolio. [documented behavior | strong] 13F cover, period 2026-06-30 information table

People, capital, and influence map

Deal-function attribution

Unknown functions remain unknown rather than defaulting to Xu. [researcher inference | strong]

Influence is clearest when a filing documents a board seat, nomination right, or signature. Interview claims that Xu recruited, coached, or pressed a strategic choice are retained as investor testimony, not proof that she caused the outcome. [researcher inference | strong]

Comprehensive deal ledger

How to read the ledger

The 44 rows below are the complete scoped public denominator—not a complete lifetime portfolio. “Round” is not “check,” “company value” is not “proceeds,” and “public holding” is not “return.” E1 means a credible filing or direct route reconstructs at least one material decision/security/governance/outcome dimension; E0 means the association is meaningful but too incomplete for economics. Every unknown is an evidence result. [researcher inference | strong]

Pre-Capital Today and Baring era — 9 paths

Capital Today era — 35 investments and admitted misses

Ex-ante forensic case studies and decision labs

The cases below freeze the investable record before revealing the outcome. “Known” means support available at or close to the decision; “believed” is the investor's retrospective reconstruction of the thesis; “disputed” flags contradictions or facts a contemporary committee could not safely assume. The questions deliberately appear before each outcome. [researcher inference | strong]

Ex-ante alternatives and dissent register

Case 1: JD.com — capital before consensus

Frozen record: January–March 2007

Underwriting view. The investable edge was not “e-commerce will grow.” It was the willingness and capacity to finance a low-trust, inventory-owning merchant through negative cash conversion while building a logistics advantage. The investment should fail if genuine-product trust does not produce repeat use, fulfillment costs do not decline with density, or reserves cannot fund the next milestone. [researcher inference | strong]

Decision lab — answer before reading the outcome

1. Would you invest then? Only with verified repeat behavior, supplier terms, inventory loss controls, and a reserve plan that treats $10m as the first—not last—capital need.
2. What would kill the deal? Negative contribution after fulfillment, counterfeit/returns eroding trust, inventory and payable stress, founder-control abuse, an unenforceable offshore/VIE structure, or an inability to raise the next round.
3. What ownership/terms fit the risk? A preferred security with information and protective rights, an employee pool sized before pricing, board representation, founder vesting/key-person protection, pro rata/follow-on capacity, and explicit related-party and VIE covenants. Exact historical terms beyond the filed securities are unknown.
4. What evidence unlocks the next check? Cohort repeat and gross profit after fulfillment; inventory turns and supplier aging; on-time delivery/return data; warehouse unit economics; management recruitment; and a financing milestone that reduces rather than compounds working-capital risk.

Outcome revealed

JD scaled, listed on Nasdaq in May 2014, and became Capital Today's best documented outlier. That does not validate every ex-ante claim: the logistics build was capital intensive, later shareholders diluted or sold, and the offshore/VIE and regulatory risks persisted. The reconstructable security path follows. [documented behavior | strong]

JD mechanical ownership and sale ledger

All share counts in the 2013 shareholder agreement are adjusted for the five-for-one split effective 2012-04-18. The filing sometimes reports aggregate syndicate transfers, so those rows cannot be assigned to Capital Today. [documented behavior | strong] JD shareholders agreement, recitals and share-adjustment clause

The 7.8% and 8.1% figures are both true snapshots: 7.8% is immediately pre-IPO and 8.1% is 2014 year-end after changes to both numerator and total shares. They must not be averaged or described as an ownership gain. [researcher inference | strong]

JD headline-number reconciliation

No audited public bridge reconciles $10m to $18m, the disclosed transfers and sales to $2.9bn, or security-level gross receipts to Fund I distributions. Missing items include pre-IPO transfer prices, every post-IPO sale, underwriting and transaction expenses, taxes, FX, recycling, LP distributions, management fees, carry, and residual holdings. Therefore none of 100x, 145x, or “more than 150x” is accepted here as audited fund MOIC. [researcher inference | strong]

Case 2: Yifeng Pharmacy — patient control without a clean exit

Frozen record: October 2008–December 2009

Decision lab — answer before reading the outcome

1. Would you invest then? Yes only if store-level cohorts, regulatory licenses, inventory controls, and a disciplined regional replication plan support the RMB430m pre-money price.
2. What would kill the deal? Reimbursement or licensing violations, counterfeit/expired stock, low same-store sales, working-capital leakage, acquisition-driven accounting, or governance conflict around a concentrated minority stake.
3. What ownership/terms fit the risk? Board representation, pre-emption and pro rata rights, audited store/cohort reporting, related-party protections, consent rights over acquisitions/debt, and a staged follow-on tied to store productivity.
4. What evidence unlocks the next check? Mature-store contribution margin, inventory days and write-offs, pharmacist/licensing compliance, new-store payback, procurement savings, and clean audit/internal controls.

Outcome revealed

Yifeng reached the public market, and Xu's director role and Capital Today Fund I's control chain are filing-supported. In 2018, a continuation transaction used roughly $250m of new capital to redeem non-rolling LP interests while other exposure continued through River entities; it was not an issuer share sale. [documented behavior | strong] Yifeng continuation filing, transaction structure

The evidence supports initial gross cost of $14.628510m + $14.646435m = $29.274945m and a 31.67% dated ownership snapshot. It does not support final proceeds or return because the later continuation price allocation, roll/cash elections, fees, carry, taxes, subsequent dividends/sales, and distributions are missing. “Patient capital” is a plausible mechanism; “successful exit” is not established by the continuation alone. [researcher inference | strong]

Case 3: Three Squirrels — online brand with channel risk

Frozen record: 2013–2014

Decision lab — answer before reading the outcome

1. Would you invest then? Only after separating genuine brand repeat from marketplace promotion and proving that supplier/food-safety systems scale with order growth.
2. What would kill the deal? Platform fee or algorithm dependence, promotion-driven low-quality revenue, food-safety failure, working-capital stress, copycat products, or an offline expansion with poor store payback.
3. What ownership/terms fit the risk? Preferred shares with pro rata and information rights, board/observer rights, food-safety and related-party covenants, founder vesting, and follow-on tranches tied to repeat/gross-margin and quality milestones.
4. What evidence unlocks the next check? Channel-level cohort retention and contribution, unaided brand demand, supplier audit failure rate, refund/complaint trends, inventory freshness, and a controlled second-channel experiment.

Outcome revealed

The prospectus documents total Capital Today cost of $5.000000m + $0.415000m + $7.430307m = $12.845307m, 18.64% ownership before the IPO and 16.73% under the stated post-offering assumption, plus Xu's directorship from December 2015. [documented behavior | strong] Three Squirrels prospectus, shareholder and director sections

The IPO created a price and liquidity path but not a documented realization. The source does not provide Capital Today's post-IPO sales, dividends, fees, carry, distributions, or residual mark. The case is valuable because primary cost and ownership are visible while the decisive brand-versus-channel question remains independent of the listing outcome. [researcher inference | strong]

Case 4: Dianping to Meituan — competition, merger, and concentration

Frozen record: before the October 2015 combination

Underwriting view. The decision was not simply which app would win. It required comparing standalone burn and differentiated supply against a combination that could reduce duplicate subsidies but introduce integration, governance, exchange-ratio, and antitrust risk. [researcher inference | strong]

Decision lab — answer before reading the outcome

1. Would you invest then? Only if cohort contribution and merchant retention remain attractive after removing subsidies, and if the merger creates measurable density rather than merely a larger loss pool.
2. What would kill the deal? Subsidy-dependent demand, merchant multi-homing with no pricing power, failed integration, founder/governance deadlock, capital exhaustion, or antitrust/data restrictions.
3. What ownership/terms fit the risk? Pro rata capacity, information and board/observer rights, merger-consent and exchange-ratio protections, preference treatment clarity, related-party controls, and explicit reserve limits for subsidy wars.
4. What evidence unlocks the next check? City-level cohort contribution, merchant retention, cross-sell adoption, overlap savings, delivery density, integration milestones, regulatory analysis, and a financing plan to cash-flow breakeven.

Outcome revealed

The combined company listed in Hong Kong. Its prospectus shows CTG Evergreen Investment XIX with 45,472,421 shares / 0.9075% and LT Growth Investment XIV with 76,466,111 shares / 1.526%. Combined: 45,472,421 + 76,466,111 = 121,938,532 shares and 0.9075% + 1.526% = 2.4335%, rounded here to 2.43% pre-offering. The filing indicates about 2.22% under its post-offering assumptions. [documented behavior | strong] Meituan prospectus, principal-shareholder table

Those snapshots do not reveal original Dianping/Meituan costs, merger conversion, interim transfers, post-listing sales, dividends, or LP distributions. The 2021 SAMR decision—RMB3.442bn fine and merchant-deposit remediation—confirms platform regulation was a material company risk; it is not evidence of misconduct by Xu or Capital Today. [contemporaneous record | strong] SAMR decision summary

Case 5: Dingdong Fresh — an ordinary public mark

Frozen record: May 2019 Series B3

Decision lab — answer before reading the outcome

1. Would you invest then? Only with mature-city contribution evidence, realistic spoilage and rider costs, and financing capacity through a downside demand case.
2. What would kill the deal? Negative contribution in mature cohorts, uncontrolled shrink, labor/regulatory shocks, subsidy-dependent retention, warehouse overbuild, or a liquidity runway shorter than the path to self-funding.
3. What ownership/terms fit the risk? Preferred downside protection, board/observer and monthly unit-economic reporting, pro rata rights, debt/capex controls, and milestone-based follow-ons rather than a single large exposure.
4. What evidence unlocks the next check? Mature fulfillment-station payback, order density and basket margin, customer cohort contribution after promotion, shrink/spoilage, rider utilization, working-capital cycle, and a verified runway.

Outcome revealed

Dingdong listed in the United States. Capital Today Evergreen Fund's Q2 2026 13F reports 11,879,332 ADS valued at $22,927,111. That is a dated U.S.-listed long-position value, not a clean comparison with the 2019 $35m private investment: security conversion, additional purchases/sales, dilution, distributions, and whether the legal holders are economically identical are not reconciled. [documented behavior | strong] 13F information table, Q2 2026

On 2026-02-05, Dingdong announced a definitive agreement to sell substantially all of its China operations to Meituan for $717 million in cash consideration to the company. Closing is subject to shareholder approval, SAMR antimonopoly clearance, completion of an overseas-business carve-out, non-resident indirect-transfer tax filings, absence of a material adverse effect, and other conditions. It is a signed but unclosed corporate transaction—not Capital Today proceeds, realization, or return. [documented behavior | strong] Dingdong Form 6-K, Exhibit 99.1, transaction consideration and closing conditions

This is intentionally an ordinary case. The public mark is neither a write-off nor proof of a successful fund outcome. An investment committee should grade the original thesis on cohort and capital-efficiency evidence, and the vehicle outcome only when a full cost/ownership/cash-flow bridge exists. [researcher inference | strong]

Case 6: Real Kungfu — governance before growth

Frozen record: 2007 entry and the reported 2010 secondary proposal

Underwriting view. The central risk was not restaurant demand. A roughly 47/47 founder split could paralyze appointments, financing, related-party decisions, and exits. A larger secondary purchase without verified title, spousal/property consents, closing conditions, escrow, and a workable shareholder agreement would magnify rather than solve that risk. [researcher inference | strong]

Decision lab — answer before reading the outcome

1. Would you invest then? Only with independent forensic diligence and a shareholder agreement that resolves the founder deadlock; otherwise the governance risk overwhelms an attractive consumer category.
2. What would kill the deal? Disputed title, undisclosed related parties, diversion of company assets, unreliable financial controls, founder deadlock, loss of licenses/food safety, or inability to enforce a secondary closing/refund.
3. What ownership/terms fit the risk? Escrowed consideration, title and spousal/property representations, closing conditions, indemnity/security, independent directors, audit and bank-account controls, related-party approval, deadlock resolution, key-person provisions, and put/remedy mechanics that are actually enforceable.
4. What evidence unlocks the next check? Beneficial-ownership and marital-property opinions, bank/related-party forensic audit, store-level reconciliation, litigation and police checks, signed shareholder consents, escrow confirmation, and independent references on both founder blocs.

Outcome revealed and legal-state controls

A Guangzhou prosecutor/court publication records that founder Cai Dabiao's relevant acts occurred in 2009–2010 and that he received a 14-year first-instance sentence on 2013-12-12 for crimes including embezzlement and misappropriation involving more than RMB30m. That is an adjudicated first-instance state concerning Cai; it is not a finding that Xu, Fang Min, or Capital Today participated in or knew of the conduct. [contemporaneous record | strong] government prosecutor/court case PDF, case facts and judgment

Separate press accounts say the contemplated 21.25% secondary did not close after Cai's arrest, Capital Today sought return of a $3.8m advance, Xu resigned from the board in September 2011, Fang Min later served, and the investment entity was transferred in late 2012. These are reported procedural/economic facts, not adopted judicial findings; the final refund, consideration, ownership, and fund loss remain unresolved. [contemporaneous record | provisional] Sina report, transaction and board chronology Guangzhou court report, shareholder litigation context

The postmortem is adverse even without a quantified loss: the public record shows the consequence of unresolved founder control, title, and enforceability. It does not support accusing the investor of the founder's crimes or claiming the whole reported RMB425m was paid/lost. [researcher inference | strong]

Portfolio-level forensics

What changed, and what persisted

The durable pattern is concentration plus reserve support, not a universally demonstrated formula. JD's bridge financing and Xingsheng's five reported follow-ons show the behavior; Xu's “top three to five ≈ 90%” statement describes a desired/remembered outcome distribution. Without a full denominator, reserve ledger, and realized attribution, the evidence cannot determine whether concentration caused the record, merely followed early marks, or amplified both winners and losses. [researcher inference | strong] Xu 36Kr interview

Eight portfolio questions

1. What was the real edge? The best-supported edge was willingness to fund capital needs and hold through uncertainty after identifying consumer trust, distribution, network density, or regional-scale economics—not access to a uniquely proprietary technology. JD, Yifeng, and Meituan support this; the conclusion remains selected-case inference. [researcher inference | moderate]
2. Where did ownership matter? Filing-visible positions—JD 7.8% pre-IPO, Yifeng 31.67% after the second round, Three Squirrels 18.64% pre-IPO, BOSS 14.7% pre-IPO, and Meituan 2.43% pre-offering—show meaningful but heterogeneous exposure. Percentages at different dates and companies are not directly comparable. [documented behavior | strong]
3. Did the team follow on? Yes in documented/reported examples: JD Series B participation, Yifeng's second investment, Three Squirrels' 2014 primary, BOSS's $20m Series E, Xingsheng's five reported follow-ons, and Moonshot's reported add-on. The portfolio-wide follow-on rate and reserve ratio are unknown. [documented behavior | moderate]
4. Was governance part of the playbook? Board roles appear in JD, Yifeng, Three Squirrels, BOSS, NetEase, and Real Kungfu. Yet a seat's practical influence, attendance, dissent, committee work, and information quality are seldom public. Real Kungfu is a warning that formal presence does not guarantee effective control. [researcher inference | strong]
5. How did liquidity occur? IPO sales and secondaries are partly visible for JD and Zhihu; Yifeng used a continuation transaction; mergers created paths for Ganji/Dianping; many private/current rows have no disposition evidence. Company liquidity must not be equated with LP DPI. [documented behavior | strong]
6. How much was founder selection versus market selection? Xu's testimony emphasizes Liu Qiangdong, Wang Xing, and Yang Haoyong, while admitted PDD/Kuaishou misses show the cost of a screen that failed to update around product/consumer behavior. The public record cannot isolate founder-picking alpha from category and capital-cycle exposure. [researcher inference | moderate]
7. What did regulation change? VIE/foreign-ownership limits shaped offshore internet structures; antitrust/data/overseas-listing actions later affected Meituan, BOSS, and peer platforms. Regulation changed permissible structures, listing timing, valuation, and operating conduct; it was not just a discount-rate input. [researcher inference | strong]
8. Can the public record reproduce the returns? No. JD permits partial gross-sale arithmetic, Yifeng and Three Squirrels permit cost/ownership arithmetic, and the 13F permits a current public-position calculation. None supplies fund-level net cash flows, and the 44-path public set is not the full portfolio. [researcher inference | strong]

Flagship, ordinary, and adverse comparison

Current concentration control

The Q2 2026 13F has just three positions, with PDD accounting for 81.6% of the reported $129.5m value. This is consistent with concentration inside that U.S.-listed long-only filing, but it cannot establish total vehicle concentration because private securities, non-U.S. listings, cash, derivatives, short positions, and other exempt assets are outside Form 13F's reported set. PDD's presence also does not reverse the earlier admitted miss: buying public ADS later is economically and decisionally different from investing before the Pinhaohuo/Pinduoduo breakout. [researcher inference | strong] Q2 2026 13F cover information table

Dingdong’s February 2026 agreement to sell substantially all China operations for $717 million remained subject to closing conditions; its corporate consideration cannot be substituted for the $22.927 million June 30 fund mark or treated as a Capital Today distribution. [documented behavior | strong] Dingdong Form 6-K, Exhibit 99.1

Failure, miss, and anti-portfolio postmortems

Adverse cases

Misses and anti-portfolio

Anti-hindsight protocol

For every pass or sale, preserve: decision date; information set; explicit objection; proposed stage/check/security/ownership; available reserves; named decision owner; next evidence trigger; and opportunity-cost benchmark. Review six, twelve, and twenty-four months later without rewriting the original memo. A missed company's later valuation is never the “loss”; the relevant quantity is the probability-weighted return of an investable allocation under the terms actually available, which is absent for these misses. [researcher inference | strong]

Uncertainty and retrieval ledger

What remains unresolved

Numerical audit rules applied

• Every computed figure shows its inputs: JD IPO gross (5,756,820 ordinary / 2) × $19, JD December secondary 3.6m ADS × $23.80, Yifeng initial cost $14.628510m + $14.646435m, Three Squirrels cost $5m + $0.415m + $7.430307m, Meituan holding 45,472,421 + 76,466,111, and 13F PDD share $105,648,105 / $129,498,857. [researcher inference | strong]
• These are gross security/value calculations as of their dated sources. None includes the complete cash flows, allocations, fees, carry, taxes, FX, recycling, distributions, or residual assets required for net fund return. [researcher inference | strong]
• Independent corroboration is strongest for transaction mechanics that appear in regulator/issuer filings. A claim repeated across profiles that ultimately trace to Xu or Capital Today is still one-source testimony, not independent triangulation. [researcher inference | strong]
• Financing totals—Guazi nearly $600m, Outer $50m Series B, and reported AI rounds—are company/syndicate amounts unless a source expressly identifies Capital Today's allocation. [researcher inference | strong]

Source ledger

Grades: A = regulator, court, exchange, issuer-filed primary document, or direct first-party announcement; B = credible contemporaneous reporting or direct interview; C = retrospective/secondary profile or lead used only with qualification. The limitation column preserves whether an A-grade item is unaudited, self-published, mirrored, or otherwise narrower than a filing.

Saturation and stopping test

Research proceeded in English and Chinese through four passes: (1) identity/vehicle/portfolio discovery; (2) filing-grade security, ownership, governance, and exit reconstruction; (3) criticism, lawsuit, fraud, antitrust, data-security, failure, miss, and anti-portfolio sweeps; and (4) exact-number contradiction and alias reconciliation. The fourth pass added Guazi's named financing route, the explicit Kuaishou/PDD miss account, the 2026 reported vehicle-transition lead, and the Q2 2026 13F; it did not produce a new audited fund-return source or a defensible full portfolio denominator. [documented behavior | strong]

A final targeted sweep of JD's $10m/$18m and $2.4bn/$2.6bn/$2.9bn claims converged on the same primary share documents and derivative profiles. A final Real Kungfu sweep converged on the government case record, court context, and press-reported transaction dispute without a recovered final fund-loss ledger. Further general web searching is therefore low-yield; the next gains require the specific nonpublic or registry artifacts in the retrieval table. [researcher inference | strong]

The dossier is saturated for public-source decision and security reconstruction, not for fund performance. It supports a rigorous conclusion narrower than the mythology: Xu led or governed several unusually consequential Capital Today positions, JD is a real but incompletely monetized public outlier, patient/concentrated follow-on behavior is documented in selected cases, and the true career return distribution remains unknowable from the public record. [researcher inference | strong]

Coverage and evidence base

The public written corpus is genuinely thin. Three writing-form records were found: one full authored investment essay, one full organizational statement reproduced by a named newsroom, and one metadata-only attributed book chapter. The two accessible priority items have notes and are analyzed below. The book chapter “今日资本：什么变了，什么没变” in 《增长的本质》 was visible only through catalog/product metadata; no full text, authorial provenance, or stable preview was available, so it is inventoried but not content-analyzed. Edited talks, press quotations, official-firm reposts, and editorial compilations are excluded from writing breadth. No public shareholder/LP letters, personal blog, newsletter, public memo archive, or long-thread corpus was found.

Source-by-source analysis

1. “巴菲特股东大会回来读段永平” (2025-05-09)

Authorship, context, and access. Huxiu labels Xu as author and says Munger Academy invited the reading note; the complete four-part text is available. It follows Berkshire's 60th annual meeting and is the only full original investment essay located. Huxiu, title/byline and introductory paragraph

Central argument and reasoning. Xu begins from “buying a stock is buying a company” and future cash flows, then asks how many companies one can understand, how much safety margin uncertainty requires, and how ethical/long-term discipline constrains action. Business quality and culture precede price, yet an investor must understand the business well enough to approximate ten-year cash flows. Huxiu, “一、价值投资的本质” and “三、安全边际” This is a hybrid rather than a conversion from VC to slow-growth value investing: she argues that technology companies can qualify if adoption has crossed a chasm and the eventual winner can be held for ten years. [investor-stated | strong]

Evidence and assumptions. Most examples—NetEase, Apple, Moutai—describe Duan Yongping's record, not Capital Today's. Her own reasoning assumes a rare great company can be recognized early enough, future cash flow can eventually be bounded, and concentrated holding survives financing, governance, dilution, regulation, and fund-liquidity constraints. Those assumptions are not tested in the essay. [researcher inference | strong]

Decision rules and implications. The actionable sequence is: remain inside a competence circle; require business/culture quality; estimate long-duration cash flow; seek a visible discount/safety margin; concentrate only when understanding is sufficient; and hold through volatility. Her home-run formulation distinguishes frequency (era/luck) from intensity (ownership and duration). Huxiu, “二、能力圈” and “三、安全边际” [investor-stated | strong]

Evolution and contradiction. Xu candidly writes that valuation had not seemed very important in VC and she did not know how to handle valuation or safety margin. Huxiu, “三、安全边际,” paragraph beginning “估值对风险投资” That sits in real tension with her 2015 advice that a $40m versus $50m entry matters little if the company succeeds. PE Daily, “错失两个半公司” The stable element is concentration/holding; the new element is explicit price discipline. [investor-stated | strong] for the change in language; actual implementation is [documented behavior | insufficient public record].

2. Capital Today four-point statement (2024-07-10)

Authorship, context, and access. The Paper reports that Xu posted a statement to WeChat Moments and reproduces four numbered points. It answers a commentary alleging closure of the primary team, LP contribution problems, and office shutdowns. The Paper, paragraphs 31–47

Central argument and evidence. The statement offers falsifiable counterexamples to a total primary-market exit: Moonshot's May 2023 first round, a July 2024 follow-on, and a May 2024 SandAI lead. It separately denies LP delays and office closures, while acknowledging normal personnel adjustment. The Paper, four numbered points [investor-stated | strong] as a statement of position; transaction and staffing facts require independent evidence.

Assumptions and limits. The rhetoric treats a handful of new investments as sufficient to refute “withdrawal,” which is fair against an absolute allegation but does not answer allocation mix, team headcount, pace, check size, or whether public-market research became dominant. [researcher inference | strong] The later Q2 2026 13F establishes meaningful public holdings but not abandonment of primary markets. SEC 13F information table

Operational implications. Use the statement as an entity/activity checkpoint, not philosophy. DEALS should verify Moonshot/SandAI rounds, identify the investing vehicle and team, and compare primary deployment with the disclosed public portfolio. [researcher inference | strong]

Cross-source synthesis

• Long duration survived; valuation language changed. The 2025 essay's concentrated long-hold logic matches the firm's longstanding rhetoric, but adds an explicit admission of a valuation skill gap and a DCF/safety-margin remedy. [investor-stated | strong]
• The vehicle is part of the doctrine. The 2024 statement stresses that the evergreen vehicle was already fully funded; the 2025 essay emphasizes holding duration. Long-duration capital plausibly enables patience, but compensation, liquidity, and succession terms remain unknown. [researcher inference | moderate]
• Writing scarcity matters. The polished philosophy is mostly spoken and retrospectively edited. One essay cannot establish how frequently rules were used, whether they predated outcomes, or whether dissenting internal memos existed. [researcher inference | strong]

Intellectual chronology, evolution, and contradictions

Decision playbook

Modern VC translation (as of 2026-08-23)

• Transfers: explicit competence boundaries, scenario cash flows, ownership/holding analysis, and separating market-cycle luck from controllable decision intensity. [researcher inference | strong]
• Does not transfer mechanically: a 28-year/evergreen vehicle's patience cannot be assumed in a ten-year fund; China's 2006–2018 platform economics and financing/regulatory environment do not establish 2026 AI outcomes. [researcher inference | strong]
• Observable applicability signals: repeatable unit economics, proprietary feedback/data loops, low dependence on perpetual subsidy, governance that permits patient reinvestment, and fund reserves consistent with claimed duration. [researcher inference | moderate]
• Likely misuse: invoking “great company” to waive price, interpreting a high-growth story as a DCF, or using concentrated winners without showing the miss denominator. [researcher inference | strong]
• Grounded current case: Xu's 2026 podcast applies the same category-map and concentration language to embodied AI, while warning that software suppliers without data cannot create a flywheel. Because only chapter metadata was processed, this is a retrieval target rather than a complete application proof. 高能量, 15:06 and 1:38:18 chapter locators [investor-stated | provisional]

Best writings and unresolved record

1. “巴菲特股东大会回来读段永平” — the only full authored investment essay and the clearest evidence of changed valuation language.
2. 2024 four-point statement — valuable for bounded organizational facts, not philosophy.
3. 《增长的本质》 chapter — third only as a retrieval lead; metadata cannot support content claims.

There are not five defensible writings. Adding edited talks or recycled compilations would manufacture a corpus. The highest-value unresolved work is obtaining the book chapter and any legitimately public internal/LP letters or memos; no such letters were found through firm, web, archive, catalog, and bilingual query routes.

Coverage and evidence base

Eighteen worthwhile underlying appearances were identified after collapsing mirrors, translations, and syndicated copies. Eleven have full/near-full edited text, three have substantive excerpts or official summaries, two have accessible full audio but no public transcript processed, and two are metadata-only. Every ★★★/★★ item in sources.md has a note; the 2016 SuperReturn panel and 2018 WSJ conversation remain ★ metadata leads because no recording or transcript was found.

The 36Kr/KrASIA pair, NJU/21st Century pair, Munger/Sina/Capital Today pair, 2016 BOSS mirrors, and Capital Allocators/legacy-firm pair each count once. Chinese is the controlling language for Chinese talks; English translations are convenience copies and do not override ambiguous names, dates, or technical terms.

Dated source-by-source guide

2014 Nanjing University commencement address

The official full transcript is strongest for biography and self-conception: three consequential choices, then passion, focus, and role models. NJU, full address It is retrospective and motivational; its “2009” ChinaHR date is contradicted by contemporaneous pre-2005 records and should remain a visible transcript error. [investor-stated | strong] for the beliefs; chronology at that line is [investor-stated | insufficient public record].

2014 “不第一，毋宁死”

The edited speech links category creation, first-mover advantage, aggressive financing, and service/logistics to JD and NetEase. TMTPost, JD and platform sections Its October 2004 JD meeting date conflicts with the more repeated 2006 account and 2007 issuance; use its doctrine, not that date, without filing reconciliation. [investor-stated | strong] for doctrine; [contemporaneous record | strong] for the conflict.

2014 Wharton Global Forum

The official event wrap-up preserves “He says there's only number one, there's no number two,” demonstrating that category concentration was public rhetoric by 2014. Wharton PDF, forum coverage section The excerpt is too short to prove a universal rule. [investor-stated | moderate]

2015 “创始人该把时间花在哪儿?”

This full transcript is unusually valuable because it names misses—7 Days, Deppon, half-credit Vancl—and an annual three-day review process. It also articulates speed, user growth, option pools, repeat purchase, and direct founder attention. PE Daily, headings “错失两个半公司” through founder-time section The miss lessons are retrospective; they show learning claims, not counterfactual certainty. [investor-stated | strong]

2015 HKEX Women Exchange luncheon

The two-page official summary adds a China-transition setting, a first-mover maxim, and performance-management language. HKEX PDF, pp.1–2 It is useful as a dated excerpt, not a substitute for the lost Q&A. [investor-stated | moderate]

2015 Chaos Academy

The full edited talk supplies the most explicit early governance doctrine: investor as accelerator; enough capital to build leadership; meaningful employee options; no earnout/redemption structure that lets investors avoid entrepreneurial risk. Chaos Academy, JD, platform, and governance sections [investor-stated | strong] The broad 7-2-1 market claim needs category-specific testing.

2016 BOSS financing event

The “capital winter” talk qualifies blitzscaling with liquidity discipline: roughly 18 months of runway, fundraising before necessity, repeat purchase, and customer-acquisition payback within two years (three at most). 36Kr, runway and unit-economics sections [investor-stated | strong] This is a genuine boundary condition on the 2015 “run at all costs” rhetoric.

2017 Guazi event

The full article attributes Ganji marketing and COO recruitment help to Capital Today and prefers product-led founders with high learning speed. Sohu event page The article body was revalidated on 2026-08-23; its edited-speech format still limits verbatim certainty. [investor-stated | moderate]

2018 two-hour 36Kr interview

This is the richest pre-2020 method source. Xu describes winner-pattern research across global retailers, direct store/customer observation, founder integrity and learning tests, and errors on ByteDance, Kuaishou, and Pinduoduo. Complete mirror, winner and miss sections It also offers a testable false-positive admission: a liked company lost because its competitor was stronger. [investor-stated | strong] Later Yonghui losses/closures weaken the interview's near-term success inference. The Paper, Yonghui outcome

2018 Yicai interview

Xu uses capital committed—not praise—as evidence of Meituan conviction and says special approval was needed after reported Ganji proceeds rolled into the position. Yicai, first three Q&A blocks She pairs Wang Xing's learning/integrity/thrift with a large, high-frequency, essential market. [investor-stated | strong] The article does not disclose check, security, ownership, or realized economics.

2020 Ray Dalio panel

The partial report shows post-miss diligence: roughly 2,000 pages of Huang Zheng material, founder speeches, and customer interviews. Sina, Xu/Pinduoduo segment No complete transcript/timestamps were available; attribution beyond the reported segment is not made. [investor-stated | moderate]

2022 Capital Allocators Ep.243

The publisher's page verifies a 2022-03-28, 53-minute episode covering biography, sourcing, founder help, evergreen structure, regulation, and competition. Capital Allocators, episode metadata and description Full audio is accessible, but the transcript is premium and was not used. The firm's legacy “May 2018” label is wrong. No content claim here exceeds the public topic map. [researcher inference | insufficient public record] pending timestamped transcription.

2022 NJU alumni interview

Xu identifies two mistakes—selling great companies too early and seeing trees without the forest—then states the three-method workflow and 2,000 annual interview program. Full mirror, opening/method and household sections She says three to five deals produced over 90% of returns, but names no fund, period, cash flows, or gross/net basis. The claim is reported, not a computed portfolio result. [investor-stated | strong] for what she said; economic support is [investor-stated | insufficient public record].

2023 Munger Academy event

The 39-statement digest makes her synthesis explicit: VC invests in change; value investing in what does not change; growth attracts her more than cheapness; only a few major opportunities are required; ownership and duration control home-run intensity. Smart Investor/Sina, items 1–17 The edited selection is not the whole talk. [investor-stated | strong]

2025 “我的投资方法论”

This full edited speech is the clearest modern process statement. It explains how missing ByteDance moved winner research into the morning schedule and how the Pinduoduo miss led to user interviews and household stays. Sohu/Forbes-Munger transcript, ByteDance and PDD sections Her “four great companies,” 20% best-case hit rate, and six-to-eight-company attention limit are valuable hypotheses but self-reported heuristics, not measured base rates. [investor-stated | strong] for the process; [investor-stated | moderate] for the numerical heuristic.

2026 高能量 Vol.227

The 2:32:36 audio is the best current route. Apple's chapter index locates category mapping (02:03), OpenAI/Anthropic (11:39), autonomous-driving data flywheels (15:06), blitzscaling boundaries (31:30), ByteDance/Meituan concentration (40:13), financing cycles (1:23:42), home-run duration (1:38:18), opportunity-cost risk (1:58:36), and AI budget discipline (2:28:01). Apple Podcasts, chapter index The audio was not transcribed; editorial chapters are not quoted as Xu's verbatim words. [investor-stated | provisional]

Recurring stories and frameworks

Three-part research compression

The mature formulation is: study historical winners, conduct direct consumer insight, and inspect every serious company in the category. 2025 methodology talk, three-method section The 2008 profile records a three-to-six-month, half-dozen-company category sweep, showing an early precursor; the 2018 interview adds annual reports, biographies, videos, and store observation. Forbes 2008, process paragraphs; 36Kr interview mirror, winner-pattern section [investor-stated | strong] for continuity.

Category leadership plus capital intensity

From 2014 onward Xu repeatedly argues that platform categories concentrate and that insufficient financing can forfeit first-mover advantage. TMTPost, category/JD sections; Chaos Academy, price-war section Meituan shows the follow-on version: reported proceeds from one winner re-concentrated into a later platform. Yicai, Meituan answers [investor-stated | strong]; efficacy across the whole portfolio remains [researcher inference | insufficient public record].

Founder trust with operating tests

Integrity, learning speed, thrift, product instinct, and willingness to share equity recur. Yet Xu also says a strong founder can lose to a stronger business/competitor and uses repeat purchase/CAC payback as checks. Yicai, Wang Xing answers; 2018 36Kr interview, false-positive section; 2016 BOSS talk, unit-economics section [investor-stated | strong].

Long duration and home-run intensity

Later talks turn “hold longer” into a formal distinction: opportunity frequency is partly era/luck; ownership and duration determine intensity. 2022 NJU interview, home-run section; 2023 Munger talk, items 11–17 The official fund page confirms long-duration vehicle tranches, but not net performance. Capital Today, “Our Fund” [investor-stated | strong] for the rule; causal vehicle advantage is [researcher inference | moderate].

Evolution, contradictions, and retrospective bias

Deal rationales and attribution boundaries

• JD: no-advertising monthly growth, integrity, category expansion, logistics/service, and option-pool recruitment recur. Exact date/check/return must come from filings, not speeches. Xu narrates the decision; other team roles remain unknown.
• Meituan: founder learning/integrity/thrift plus large frequent essential demand; the economically important action was a special-approved post-merger concentration. Do not call it an early-originated Xu deal.
• Ganji/Guazi: marketing/COO operating support is claimed; Guazi appears to have been declined early and entered later, making it a useful boundary case.
• Pinduoduo: initially missed because Xu/team judged through affluent-user experience; later research and a large disclosed public position are documented separately. Do not invent purchase thesis/timing.
• Yonghui Life/Yunchuang: rapid operating metrics influenced enthusiasm, but later losses/closures require a full decision lab and entity reconciliation.

Listen/read these five first

1. 2025 “我的投资方法论” — most complete current method, misses, and process changes; full edited text.
2. 2018 36Kr interview — deepest pre-change method record and unusually candid misses/false positive.
3. 2015 Chaos Academy — clearest early governance and capital-acceleration doctrine.
4. 2022 NJU alumni interview — mature three-method framework, consumer immersion, and holding-period lesson.
5. 2026 高能量 Vol.227 — best current AI application, but only after repository-approved transcription and speaker separation.

The 2023 Munger talk is the next-best substitute when audio transcription is unavailable; it is stronger for concentration/temperament than for operating method.

Modern VC translation (as of 2026-08-23)

• Build a category map before an IC decision: global precedents, every credible current entrant, direct users, and a written “forest” view. [researcher inference | strong]
• Treat speed as conditional: accelerate only with validated retention/repeat use, defensible acquisition payback, and enough runway to survive a funding closure. [researcher inference | strong]
• Record the exact reason for every pass and revisit misses annually; otherwise retrospective stories cannot show process improvement. [researcher inference | strong]
• Express conviction through an ownership/reserve plan tied to milestones, not adjectives. A long hold must fit the legal vehicle and liquidity promises. [researcher inference | strong]
• For AI, require a proprietary data/feedback loop and calculate how many training iterations the budget supports; the 2026 talk is a grounded lead, but the full rule awaits transcription. [investor-stated | provisional] 高能量, 15:06 and 2:28:01

Unresolved spoken record

• Transcribe Capital Allocators Ep.243 and 高能量 Vol.227 with the repository-approved OpenAI route; retain timestamped short excerpts only.
• Recover the 2018 WSJ discussion and 2016 SuperReturn panel, or document archival dead ends.
• Recover original 36Kr/NJU/Guazi recordings to test edited-transcript fidelity.
• Reconcile JD chronology and every reported return number with filings before preserving any spoken multiple.
• Treat repetition as consistency, not truth: no talk denominator can substitute for portfolio and fund cash flows.

Model 1 — Map the forest before choosing the tree

• Rule — Build a dated map of the category—customers, incumbents, startups, substitutes, historical winners, failure modes, and likely consolidators—before allowing one impressive company to dominate attention.
• Evidence status — [investor-stated | strong] as a repeated rule; [researcher inference | provisional] as a source of superior returns because no complete pass or outcome denominator is public.
• Provenance and origin — Xu's 2015 account explicitly contrasts seeing an attractive tree with understanding the forest, while later interviews describe studying the winning pattern in earlier category leaders. 2015 speech transcript, category-error discussion 2018 long-form interview, historical pattern study
• Mechanism and reasoning — A category map reduces availability bias, reveals substitutes and consolidation paths, and forces the investor to explain why this company's route to leadership is different. It also creates a denominator of alternatives against which the chosen company can be judged.
• Historical and vehicle context — The model fit China's 2005–2018 consumer and internet markets, where rapid formalization, mobile distribution, and scarce early growth capital could produce dominant platforms. A concentrated fund needed a way to choose one or two category exposures rather than own an index.
• When to use — Use it before first meetings, when a sector becomes fashionable, when a founder claims “winner takes all,” and before a merger turns two old companies into a new underwriting proposition.
• How to apply it — Write every serious candidate and substitute; score demand frequency, unit economics, product/supply control, distribution, mindshare, capitalization, regulation, and founder evidence; preserve the pass reason and the fact that would trigger re-entry.
• Examples —
  • JD.com — researcher-applied illustration. [researcher inference | moderate] Logistics and customer experience differentiated an e-commerce contender, but the category thesis must be reconstructed from filings rather than the later return legend. JD.com IPO prospectus
  • Meituan — researcher-applied illustration. [researcher inference | moderate] The relevant Capital Today decision followed the Meituan–Dianping combination, so the post-merger category map matters. Meituan IPO prospectus
  • Pinduoduo — investor-stated counterexample. [investor-stated | strong] Xu says she missed the private opportunity because she did not understand the lower-tier consumer; the favored map was incomplete. 2022 alumni interview
• Evidence — The principle appears across independently published 2015 and 2018 accounts; JD.com and Meituan primary filings anchor company and ownership facts; the Pinduoduo miss is Xu's retrospective testimony rather than a recovered pass memo. 2015 speech 2018 interview JD.com prospectus Meituan prospectus
• Limits, failure modes, and exception ledger — Category maps can favor visible incumbents, turn analogy into destiny, and miss new demand or distribution. Preserve a “map-breaker” column: different customer, lower cost, new interface, new regulation, or non-consumption. Pinduoduo and ByteDance are mandatory exceptions; a beautiful company in a weak forest is the opposite error.
• Evolution and contradictions — The early doctrine celebrates recognizing the eventual leader; later accounts foreground misses caused by an incomplete map and a narrower competence circle. This is a useful correction, but retrospective winner stories still outnumber contemporaneous losing maps.
• Modern VC translation (as of 2026-08-23) — Transfer: map foundation models, open-source substitutes, inference costs, data and workflow rights, distribution owners, regulation, and bundling. Changed condition: model components can commoditize quickly. Signal: the company retains users and rights when a platform bundles its feature. Misuse: naming a category leader before mapping substitutes. Hypothetical: reject an AI-copilot thesis that vanishes after an incumbent adds the same function. [researcher inference | strong]

Model 2 — Earn the competence circle through consumer immersion

• Rule — Treat direct observation as a way to earn and expand competence: talk to users across segments, observe real workflows or households, and state what evidence changed the investment belief.
• Evidence status — [investor-stated | strong] for the method and reported Pinduoduo learning response; [retrospective witness | moderate] for the reported scale of interviews; [researcher inference | provisional] for causal investment benefit.
• Provenance and origin — Consumer inquiry appears in Xu's 2015 channel-and-consumer framework. The later Pinduoduo account says she responded to the miss with roughly 2,000 interviews and household immersion; the count is self-reported and not independently audited. Chaos Academy transcript 2022 alumni interview
• Mechanism and reasoning — Immersion can reveal jobs-to-be-done, status, trust, price sensitivity, and channel behavior that dashboards flatten. The competence gain comes only when observations are sampled, recorded, and connected to a forecast or decision.
• Historical and vehicle context — China's regional, income, and urban/rural variation made elite-investor intuitions especially fragile. A China consumer fund could justify fieldwork as core diligence; it still needed sampling controls and a portfolio-sized learning budget.
• When to use — Use it when the investor is not a representative user, when reported demand varies by region or income, when social meaning drives adoption, or when a pass reveals a comprehension gap.
• How to apply it — Pre-register segments and questions; recruit through more than management; record positive and negative cases; triangulate interviews with retention, cohort, willingness-to-pay, and contribution-margin data; name the belief and valuation input that changed.
• Examples —
  • Xingsheng — investor-stated application. [investor-stated | provisional] Xu links household observation to community commerce and reports multiple follow-ons, although complete securities and outcomes are absent. 2022 alumni interview
  • Pinduoduo — investor-stated miss plus documented later boundary. [investor-stated | strong] The early miss is the adverse case; a later public holding is a new decision, not retroactive success. 2022 alumni interview Capital Today Q2 2026 13F table
  • Yifeng Pharmacy — researcher-applied illustration. [researcher inference | moderate] Frequent essential demand provides a plausible fieldwork setting, but the filing documents ownership rather than consumer-research causality. Yifeng prospectus
• Evidence — Independent source families establish the doctrine, the later self-reported learning loop, a public PDD position, and Yifeng's transaction facts. None exposes the field notes or an ex-ante forecast. Chaos Academy transcript Nanjing University alumni interview SEC 13F CNINFO Yifeng filing
• Limits, failure modes, and exception ledger — Convenience samples, leading questions, founder-selected users, observer effects, and vivid anecdotes can manufacture certainty. Interviews cannot replace behavior or economics. Log segment gaps, refusals, contradictory answers, and what would falsify the interpretation.
• Evolution and contradictions — The method evolves from generalized “consumer insight” into an explicit repair for class and geography blind spots. Yet Xu's later PDD holding does not disclose basis, date, vehicle, or whether immersion caused the purchase.
• Modern VC translation (as of 2026-08-23) — Transfer: shadow enterprise users through a full workflow or sample consumer users across age, language, income, and trust. Changed condition: synthetic demos and founder-selected power users can mask permission and error failures. Signal: segment evidence changes retention or willingness-to-pay assumptions. Misuse: treating interview volume as validation. Hypothetical: pause an AI-copilot check when nontechnical users cannot verify outputs. [researcher inference | strong]

Model 3 — Underwrite founder learning and integrity as behaviors

• Rule — Back founders who tell the truth under pressure, learn faster than the problem changes, recruit people stronger than themselves, and keep commitments; verify each trait through behavior and controls.
• Evidence status — [investor-stated | strong] as a selection doctrine; [documented behavior | moderate] in selected board relationships; [researcher inference | provisional] as an independent performance factor.
• Provenance and origin — Integrity, learning, and talent attraction recur across Xu's direct-voice record, including her long-form case discussions. 2018 interview, founder criteria 2025 methodology account
• Mechanism and reasoning — Private-company evidence is incomplete and conditions change. A learning founder updates the plan; an honest founder preserves data quality; a talent magnet increases organizational capacity. Governance converts those qualities from impressions into observable, enforceable behavior.
• Historical and vehicle context — Founder-led Chinese companies often faced scarce professional management, fast growth, related-party complexity, and shifting regulation. A meaningful owner and director could help recruit and govern, but familiarity and founder charisma could also weaken challenge.
• When to use — Use it at founder reference, during diligence on a metric discrepancy, before a large follow-on, after a regulatory warning, and when a board must choose between founder autonomy and additional controls.
• How to apply it — Ask for a serious error, an unpopular truth disclosed early, a senior hire better than the founder, and a changed decision; independently verify customers, employees, cap table, related parties, cash, and legal history; prewrite escalation rights.
• Examples —
  • JD.com — investor-stated application with filed ownership. [retrospective witness | moderate] Xu’s account emphasizes Liu Qiangdong’s customer commitment and persistence; filings establish ownership, not the causal founder score. 2018 interview JD.com prospectus
  • Meituan — investor-stated application. [investor-stated | strong] Wang Xing is presented as a learning founder, while post-merger success depended on a broad team and regulatory environment. Yicai 2018 Meituan interview
  • Zhen Kungfu — documented counterexample. [documented behavior | strong] The founder’s later conviction makes integrity controls—not personal confidence—the required lesson; it does not establish investor misconduct. government case record Guangzhou court report
• Evidence — Xu states the doctrine in two separately published accounts; primary company and court records anchor roles and adverse facts. No public reference file or founder scorecard supports classifier accuracy. 2018 interview 2025 methodology account JD.com SEC filing government case record Guangzhou court, shareholder context
• Limits, failure modes, and exception ledger — Charisma can impersonate learning; loyalty can conceal weak recruiting; and integrity language can become unfalsifiable. Preserve negative references, metric reproductions, related-party tests, dissent, and the point at which a governance failure changes support.
• Evolution and contradictions — The founder criterion is stable, but adverse cases make the later, control-based interpretation stricter. “Treat the founder like family” is not an acceptable governance rule; behavioral evidence and legal rights are.
• Modern VC translation (as of 2026-08-23) — Transfer: test candor and learning through observed responses to security, evaluation, privacy, and regulatory errors. Changed condition: opaque training data and benchmark contamination create new ways to hide risk. Signal: the founder discloses a material failure before diligence finds it. Misuse: treating charisma as integrity. Hypothetical: stop an AI investment when management conceals customer-data use despite exceptional growth. [researcher inference | strong]

Model 4 — Capitalize a plausible leader through evidence gates

• Rule — When category structure, unit economics, and founder quality support a plausible leader, provide enough capital to reach the next defensible position—but release follow-ons only against prewritten evidence gates.
• Evidence status — [investor-stated | strong] for aggressive capitalization and financing-winter support; [documented behavior | moderate] for repeated financing in selected cases; [researcher inference | provisional] for firm-wide effectiveness.
• Provenance and origin — Xu's 2015 speech emphasizes moving quickly and supporting the strongest companies; JD.com and BOSS narratives provide specific runway and payback illustrations. 2015 speech 36Kr BOSS account
• Mechanism and reasoning — Adequate capital can fund logistics, talent, product, and brand before competitors, while a financing winter may improve relative position. Gates prevent the same logic from becoming subsidy dependence or sunk-cost escalation.
• Historical and vehicle context — Scarce China growth capital made a committed investor valuable in 2007–2009. Later abundant capital and platform competition changed the risk: too much money could fund uneconomic expansion. Vehicle reserves and cross-fund allocation determine whether support is repeatable.
• When to use — Use it when capital purchases a measurable capability or survival interval, when competitors face the same financing constraint, and before every insider bridge or large follow-on.
• How to apply it — State check, security, ownership, dilution, runway, milestone, next financing, cumulative exposure, downside recovery, syndicate capacity, regulatory gate, and abandon rule; rerun the case with no external capital for 24 months.
• Examples —
  • JD.com — documented financing with retrospective mechanism. [retrospective witness | moderate] Filings document multiple Capital Today-related holdings; Xu’s account says financing through the crisis mattered, but complete cash flows are absent. JD.com prospectus
  • BOSS — investor-stated application. [investor-stated | moderate] A public account says preserving roughly 18 months of runway and keeping customer-acquisition payback within two years helped readiness; causality is retrospective. 36Kr BOSS account
  • Community commerce — documented regime counterexample. [documented behavior | moderate] Rapid capitalization faced subsidy, loss, and regulatory risk, showing that speed cannot repair weak contribution economics. SAMR Meituan action
• Evidence — The rule is supported by a direct speech and separately published BOSS account; SEC/HKEX and regulator records independently anchor financings, ownership, and adverse regime facts. 2015 speech BOSS shareholders agreement JD.com SEC filing SAMR
• Limits, failure modes, and exception ledger — Capital can conceal low retention, bad working capital, weak governance, or regulatory illegality. “Leader” may be defined by fundraising rather than customers. Log every gate miss, insider-only round, valuation reset, regulatory dependency, and support decision after the original thesis breaks.
• Evolution and contradictions — Earlier rhetoric makes speed and home-run asymmetry sound price-insensitive; later work emphasizes payback, competence, DCF, and margin of safety. The disciplined synthesis requires both conviction and staged evidence.
• Modern VC translation (as of 2026-08-23) — Transfer: fund proprietary workflow, distribution, evaluation, security, and unit-cost progress through staged gates. Changed condition: falling model prices and platform bundling can erase compute-led advantage. Signal: retention and payback improve after each tranche. Misuse: capitalizing benchmark theater. Hypothetical: release an AI company’s next tranche only after vendor-independent gross-margin progress. [researcher inference | strong]

Model 5 — Seek control of product, channel, and mindshare

• Rule — Prefer companies that can defend three linked controls: distinctive product or supply, reliable customer access, and trusted mindshare or brand.
• Evidence status — [investor-stated | strong] as an operating framework; [documented behavior | moderate] in selected consumer cases; [researcher inference | provisional] as a universal moat model.
• Provenance and origin — Xu's Chaos Academy account describes consumer and channel logic and the importance of brand, while Capital Today presents discipline and company-building as firm traits. Chaos Academy transcript Capital Today
• Mechanism and reasoning — Product or supply earns use, channel captures demand economically, and mindshare lowers search and trust costs. Weakness in one layer lets suppliers, platforms, or competitors extract value from the others.
• Historical and vehicle context — The frame emerged in China's retail and marketplace expansion, where store networks, logistics, mobile traffic, and consumer brands competed for control. It is most relevant to operating-heavy, concentrated investments where the investor can help recruit or measure.
• When to use — Use it for consumer, marketplace, healthcare retail, and workflow businesses; before paying a leadership premium; and when growth depends on a third-party platform or subsidized acquisition.
• How to apply it — Score each control separately; identify legal ownership and switching costs; measure gross margin, repeat, payback, organic acquisition, supplier concentration, and brand trust; run a loss-of-channel and price-war scenario.
• Examples —
  • JD.com — researcher-applied illustration. [researcher inference | moderate] Logistics and customer experience support product/channel control, while SEC filings anchor company history rather than the reconstructed moat. JD.com prospectus
  • Three Squirrels — researcher-applied illustration. [researcher inference | moderate] Brand plus online distribution created scale, but the filing cannot prove durable consumer mindshare or investor causality. Three Squirrels prospectus
  • Yonghui Yunchuang — researcher-applied counterexample. [researcher inference | moderate] Store contraction and losses show that omnichannel rhetoric without durable unit economics is not control. The Paper, Yonghui reporting
• Evidence — Direct doctrine and official firm material are independent of the primary company filings and later adverse reporting; together they support the model's content and case facts, not a measured moat score. Chaos Academy Capital Today Three Squirrels CNINFO filing The Paper
• Limits, failure modes, and exception ledger — Brand awareness is not pricing power; distribution can be rented; vertical integration can destroy returns; and supply control can invite regulation. Log platform dependence, promotional acquisition, returns, inventory, service failures, and channel conflicts.
• Evolution and contradictions — Early consumer doctrine can sound like growth plus branding. The later margin-of-safety frame requires translating “control” into cash generation and downside resilience, not reach alone.
• Modern VC translation (as of 2026-08-23) — Transfer: test control of workflow and evaluation, repeatable distribution, and trusted outcomes. Changed condition: model APIs and audience reach are rented and can be repriced or bundled. Signal: retention and economics persist after a vendor switch. Misuse: calling followers or API access a moat. Hypothetical: an AI workflow survives replacement of its foundation model without losing customer data rights. [researcher inference | strong]

Model 6 — Concentrate for home runs within an attention budget

• Rule — Concentrate capital and partner time in the few opportunities capable of driving fund-level outcomes, subject to explicit loss, correlation, reserve, and attention budgets.
• Evidence status — [investor-stated | strong] for home-run logic and a reported six-to-eight-company capacity; [documented behavior | strong] for concentration in one 13F snapshot; [researcher inference | provisional] for the complete portfolio or returns.
• Provenance and origin — Xu has repeatedly described top positions dominating outcomes and, in a later account, a personal capacity of roughly six to eight companies. 2023 public remarks 2025 methodology account
• Mechanism and reasoning — Venture outcomes are skewed, while partner attention is finite. Concentration makes correct insight consequential and gives board work enough time; explicit budgets keep a single thesis or regime from endangering the vehicle.
• Historical and vehicle context — Capital Today's focused strategy and later evergreen tranches could support larger holds and follow-ons. The public 13F shows only reportable U.S.-listed long equity and cannot establish total fund concentration, private exposure, or cost.
• When to use — Use it after a complete category comparison, when ownership-adjusted upside matters at fund scale, and at every follow-on or public-holding decision.
• How to apply it — Set maximum loss, factor correlation, ownership target, reserves, partner-hours, and liquidity constraints; compute the fund with the top position at zero, at half value, and requiring double the planned follow-on.
• Examples —
  • JD.com — documented ownership with reported economics. [documented behavior | strong] A large visible ownership became economically plausible as a driver, but the reported 150–160x lacks a public cash-flow bridge. JD.com Schedule 13G
  • PDD — documented current concentration boundary. [documented behavior | strong] It represented about four-fifths of the Q2 2026 disclosed U.S.-listed long-equity value; the filing says nothing about the full fund. Capital Today Q2 2026 13F table
  • ByteDance — investor-stated counterexample. [investor-stated | moderate] A missed outlier shows the opportunity cost of a competence boundary while demonstrating that concentration cannot cover every winner. 2025 methodology account
• Evidence — Two direct-voice sources establish the rule and capacity claim; SEC filings independently establish JD.com and current public-position snapshots. No audited vehicle statement provides a literal top-position contribution. 2023 remarks 2025 account JD.com 13G 13F cover
• Limits, failure modes, and exception ledger — Marking concentration is not cash-return concentration; public filings omit private assets and realized positions; correlated China regulatory or consumer exposure can defeat company-level diversity. Log ownership, basis, liquidity, currency, regulation, and attention debt.
• Evolution and contradictions — Concentration is stable, while late value language adds price and downside. The unresolved contradiction is that the strongest concentration claims coexist with the weakest complete denominator.
• Modern VC translation (as of 2026-08-23) — Transfer: budget loss, reserves, factor correlation, and partner attention together. Changed condition: nominally different AI companies may share the same model vendor, compute chain, regulation, customer budget, or distribution platform. Signal: the fund survives a zero on its largest position. Misuse: treating the 13F as the whole portfolio; PDD was 81.6% of only the Q2 2026 disclosed U.S.-listed long-equity value. [researcher inference | strong] SEC 13F cover SEC table

Model 7 — Finance through winters with prewritten milestones

• Rule — Reserve capital and operational attention so a high-conviction company can survive a financing winter, but write the evidence, exposure, and abandonment gates before the winter arrives.
• Evidence status — [investor-stated | strong] for the financing-winter doctrine; [documented behavior | moderate] for selected repeat investments; [researcher inference | provisional] for whether support improved aggregate returns.
• Provenance and origin — JD.com is Xu's canonical financing-winter story; BOSS and Xingsheng accounts extend the practice to cash conservation and repeated follow-ons. Chaos Academy transcript, JD.com financing 2022 alumni interview, Xingsheng follow-ons
• Mechanism and reasoning — Scarce external capital can remove competitors and make committed reserves disproportionately valuable. A milestone contract distinguishes valuable endurance from paying to postpone recognition of a broken thesis.
• Historical and vehicle context — The 2008–2009 crisis made runway a strategic asset. Later evergreen tranches may increase duration, but public marketing does not disclose reserve ratios, cross-fund rights, or LP liquidity. Capital Today
• When to use — Use it before the initial check, during a market shutdown, after a down round, or when insiders are the only available capital.
• How to apply it — Prewrite monthly burn, runway, contribution margin, retention, acquisition payback, hiring, regulatory, governance, next-round, cumulative-exposure, and stop milestones; name the owner and measurement date for each.
• Examples —
  • JD.com — documented financing path plus retrospective mechanism. [retrospective witness | moderate] Repeated Capital Today-related ownership and the crisis story fit the model, but exact support cash flows remain incomplete. JD.com prospectus
  • BOSS — investor-stated application. [investor-stated | moderate] Conserving cash and controlling acquisition payback illustrate readiness before a financing shock. 36Kr BOSS account
  • Xingsheng — investor-stated counter/control. [investor-stated | provisional] Xu reports about $165 million across six investments; without security, ownership, impairment, and outcome data, repeated support could be conviction or escalation. 2022 alumni interview
• Evidence — The doctrine appears in an independent transcript and alumni interview; primary SEC material anchors JD.com and BOSS transaction/legal context, but no source provides a complete reserve ledger. Chaos Academy Nanjing University interview JD.com SEC filing BOSS shareholders agreement
• Limits, failure modes, and exception ledger — Insider support may signal adverse selection, delay shutdown, disadvantage one vehicle, or increase regulatory exposure. Log every missed milestone, valuation reset, insider-only round, cross-vehicle allocation, waiver, and thesis rewrite.
• Evolution and contradictions — Patience remains stable; the later competence and margin-of-safety vocabulary makes the necessary gating more explicit. Public accounts still celebrate the survivor more than unsupported companies.
• Modern VC translation (as of 2026-08-23) — Transfer: reserve enough capital for a 24-month closed market, conditional on dated survival milestones. Changed condition: vendor repricing, security incidents, and regulatory delays can consume runway faster than hiring. Signal: retention, contribution margin, and vendor independence improve. Misuse: treating a larger follow-on as proof. Hypothetical: withhold the next tranche if a foundation-model price increase destroys payback. [researcher inference | strong]

Model 8 — Treat holding and liquidity as a second investment decision

• Rule — At every financing, merger, tender, listing, or vehicle transfer, decide anew whether the expected ownership-adjusted return exceeds the opportunity cost and governance risk.
• Evidence status — [investor-stated | moderate] for patience and the Yonghe regret; [documented behavior | moderate] for holdings at filing dates; [researcher inference | strong] for the required second-decision control.
• Provenance and origin — Xu's public record pairs long-hold language with regret over selling Yonghe too early. Capital Today's reported evergreen tranches and later public holdings make duration central. 2015 speech Capital Today
• Mechanism and reasoning — Large company value often compounds after initial proof, but price, dilution, liquidity, governance, and vehicle obligations change. Re-underwriting prevents both premature sale and “never sell” identity.
• Historical and vehicle context — Private stakes in China internet companies crossed into offshore listings and, sometimes, reportable U.S. public holdings. Lockups, VIE risk, currency, and LP terms make company success different from distributed return.
• When to use — Use it at a priced round, merger, secondary offer, continuation transfer, IPO, lockup expiry, regulatory event, or when a public position becomes a large share of disclosed value.
• How to apply it — Rebuild current basis and ownership; model tax, lockup, liquidity, dilution, governance, regulation, currency, and opportunity cost; compare sell, hold, distribute, hedge, and follow-on choices; record decision rights.
• Examples —
  • Yonghe — investor-stated application. [investor-stated | provisional] Xu cites an early sale as a missed compounding opportunity, but public cash flows are incomplete. 2015 speech
  • Yifeng — documented application boundary. [documented behavior | strong] The continuation transaction shows why a new vehicle decision needs valuation and conflict review. Yifeng continuation filing
  • ChinaHR — researcher-applied counterexample. [researcher inference | moderate] A 2007 profile reports roughly 800x paper value. [retrospective witness | provisional] No personal cash ledger supports that figure; the later company decline makes paper value and enterprise durability separate outcomes. 2007 Sina profile
• Evidence — Direct testimony establishes the patience/exit doctrine; official company filings independently establish Yifeng transfer and public-holding snapshots. No public ledger gives all sales or distributions. 2015 speech Capital Today Yifeng CNINFO filing Capital Today 13F
• Limits, failure modes, and exception ledger — Long holding can disguise illiquidity, valuation lag, or unwillingness to admit change. Early sale can be correct for a fund even if the company later rises. Log legal holder, mandate, lockup, valuation source, sale/distribution dates, and LP constraints.
• Evolution and contradictions — “Hold the winner” and later margin-of-safety language create a healthy tension: quality is not an infinite price, and an attractive historical basis is not a current thesis.
• Modern VC translation (as of 2026-08-23) — Transfer: re-underwrite every secondary, tender, continuation transfer, and public hold. Changed condition: longer private duration and continuation conflicts make price discovery weaker. Signal: current ownership-adjusted return clears opportunity cost after liquidity and governance haircuts. Misuse: using a trophy narrative to avoid a sale. Hypothetical: require independent valuation and LP-conflict review before moving an AI asset across vehicles. [researcher inference | strong]

Model 9 — Use recruiting, metrics, and brand as bounded interventions

• Rule — Concentrate investor help where it has a testable mechanism: recruit pivotal talent, install a small operating-metric system, or sharpen durable customer positioning and brand.
• Evidence status — [investor-stated | strong] for Xu's “three weapons” framework; [documented behavior | moderate] for board and transaction roles; [researcher inference | provisional] for incremental company value.
• Provenance and origin — The 2015 Chaos Academy transcript is the clearest direct account of recruiting, KPI, and brand as investor interventions; KANZHUN's filing independently anchors Xu's board role. Chaos Academy transcript KANZHUN Form 20-F
• Mechanism and reasoning — A focused intervention can remove an organizational bottleneck without making the investor a shadow executive. Dated goals and founder corroboration allow the team to distinguish help from proximity.
• Historical and vehicle context — Fast-growing Chinese companies often needed professional executives, operating cadence, and brand investment. Concentration and board ownership made this help feasible; broad portfolios or weak governance would make the same promise unrealistic.
• When to use — Use it after naming one bottleneck, owner, metric, and time box; when the founder explicitly wants help; and before claiming board participation as value-add.
• How to apply it — Create a 90-day intervention brief: problem, counterfactual, actions, decision rights, founder owner, metric, exit date, and independent postmortem. Record candidates introduced, interviews, decisions, and outcomes—not just introductions.
• Examples —
  • BOSS — documented role with inferred intervention. [researcher inference | moderate] Xu’s directorship and reported cash/payback discussion provide a plausible metric intervention, but not proof of causal outcome. KANZHUN Form 20-F 36Kr BOSS account
  • JD.com — investor-stated application. [investor-stated | moderate] Recruiting, customer experience, and financing advice recur in Xu’s account; company execution belongs primarily to Liu and the team. Chaos Academy
  • Zhen Kungfu — documented counterexample. [documented behavior | strong] A board relationship did not prevent founder/shareholder crisis, demonstrating that advice without effective controls is insufficient. government case record Guangzhou court, shareholder context
• Evidence — One direct transcript establishes the framework, while independent SEC and court sources establish board and adverse-case facts; no founder/co-investor record provides a complete intervention counterfactual. Chaos Academy KANZHUN SEC filing JD.com SEC filing government case record Guangzhou court, shareholder context
• Limits, failure modes, and exception ledger — Investors can waste founder time, recruit for prestige, impose dashboard theater, or claim founder work. Governance authority may be too weak. Log declined advice, failed searches, metric changes, founder attribution, and cases where the investor should step back.
• Evolution and contradictions — The intervention model is stable, but adverse governance cases demand a stronger boundary between helpful coaching and enforceable oversight. A “weapon” is not an outcome.
• Modern VC translation (as of 2026-08-23) — Transfer: time-box help around one measurable recruiting, metric, security, or procurement bottleneck. Changed condition: AI companies need model-risk and data-governance expertise that a generalist board may lack. Signal: a named owner and counterfactual move. Misuse: claiming all company execution as investor value-add. Hypothetical: a 90-day security-lead search with a prewritten stop date. [researcher inference | strong]

Model 10 — Convert mistakes into a recorded decision loop

• Rule — Record the decision before the outcome, revisit it on a fixed cadence, separate process error from bad luck, and convert only repeated evidence into a revised principle.
• Evidence status — [investor-stated | strong] for recording discussions, multi-day postmortems, and named misses; [researcher inference | moderate] for the proposed scorecard; [researcher inference | provisional] for return improvement.
• Provenance and origin — Xu has described recording investment discussions and spending roughly three days each year reviewing decisions. Later accounts name ByteDance and Pinduoduo as competence errors. 2015 speech transcript 2025 methodology account
• Mechanism and reasoning — A contemporaneous record resists hindsight, reveals whether a rule actually preceded the decision, and distinguishes missing evidence, wrong model, execution variance, regime change, and luck. Repetition turns anecdotes into an operating update.
• Historical and vehicle context — Long-duration private investing produces slow and noisy feedback. Annual review is better than none, but milestone reviews and pass tracking are needed so feedback arrives before a decade-long outcome.
• When to use — Use it at every IC vote, follow-on, pass, mark change, governance event, liquidity decision, and annual portfolio review.
• How to apply it — Store thesis, probabilities, disconfirming evidence, check/ownership, milestones, kill criteria, dissent, and confidence; later score forecast calibration, process adherence, role attribution, and outcome; publish the rule change and its effective date.
• Examples —
  • Pinduoduo — investor-stated miss and later documented boundary. [investor-stated | strong] The miss produced a stated consumer-immersion response; later public ownership is a separate test. 2022 alumni interview
  • ByteDance — investor-stated counterexample. [investor-stated | moderate] The pass is attributed to not understanding feed advertising, exposing a technology/business-model competence gap. 2025 methodology account
  • Yonghui — researcher-applied adverse control. [researcher inference | moderate] A fashionable new-retail thesis followed by losses and closures, testing whether postmortems constrain follow-ons rather than merely improve the story. The Paper, Yonghui reporting
• Evidence — Separate 2015, 2022, and 2025 accounts support the loop and named errors; adverse reporting independently anchors a company outcome. The actual recordings and annual review files are not public. 2015 speech 2022 interview 2025 account The Paper
• Limits, failure modes, and exception ledger — A postmortem can become polished hindsight, blame shifting, or survivor worship. Missing passes and written-off companies destroy calibration. Preserve raw memos, losing cases, unchanged rules, minority dissent, and decisions where good process produced a bad outcome.
• Evolution and contradictions — The public doctrine becomes more explicit over time: from instinct and winner patterns toward competence circles, DCF, margin of safety, and structured fieldwork. Without dated internal records, the degree and timing of the change remain uncertain.
• Modern VC translation (as of 2026-08-23) — Transfer: preserve immutable forecasts and score them later. Changed condition: AI retrieval can help surface contradictions but can also rewrite history or leak confidential data. Signal: calibration improves without changing the original memo. Misuse: letting a model rationalize the outcome. Hypothetical: query an access-controlled journal for disconfirming evidence while locking the source record. [researcher inference | strong]

Model 11 — Require price protection inside the competence circle

• Rule — Estimate value only where the investor can explain the business and its key uncertainties; use scenarios and a margin of safety so an exceptional company is not an unlimited-price investment.
• Evidence status — [investor-stated | strong] for the later DCF and margin-of-safety doctrine; [investor-stated | moderate] for earlier price-insensitive rhetoric; [researcher inference | provisional] for consistent historical application or return benefit.
• Provenance and origin — A 2015 speech argues that price matters less for rare category leaders, while a 2025 essay explicitly puts DCF, scenario uncertainty, competence boundaries, and margin of safety into the method. 2015 speech, price tension 2025 Huxiu essay
• Mechanism and reasoning — A competence boundary narrows forecast error; scenarios expose which assumptions create value; a discount to conservative value protects against business, financing, dilution, and exit uncertainty. The rule turns conviction into an ownership-adjusted expected return rather than admiration.
• Historical and vehicle context — Scarce capital and rapid China consumer growth could make entry-price errors look small beside outlier outcomes. Slower exits, continuation vehicles, public residuals, and a different rate regime make basis, liquidity, and opportunity cost harder to ignore.
• When to use — Use it before every initial check, follow-on, continuation transfer, tender, or public hold, and whenever leadership rhetoric is offered as a reason to suspend valuation.
• How to apply it — Define the competence boundary; build bear, base, and bull cash-flow cases; reconcile check, fully diluted ownership, follow-ons, dilution, exit value, timing, currency, fees, and loss probability; require a margin to conservative value and name the price that changes the decision.
• Examples —
  • Yifeng Pharmacy — researcher-applied illustration. [researcher inference | strong] Filed 2008 and 2009 subscription amounts permit a reproducible cost basis, but no complete exit cash flow supports a realized multiple. Yifeng prospectus
  • Three Squirrels — researcher-applied illustration. [researcher inference | strong] Filed subscription costs and ownership make price analysis possible; durability and eventual distributions remain separate questions. Three Squirrels prospectus
  • JD.com — investor-stated counterexample with documented boundary. [researcher inference | strong] Earlier price rhetoric and unreconciled reported value/cost multiples show why a famous winner cannot substitute for a cash-flow bridge or ex-ante valuation record. 2015 speech JD.com return reconstruction
• Evidence — Two independent direct-voice publication families establish the evolution in doctrine, while issuer filings independently anchor reproducible company costs. No public IC memo shows that the later valuation rule governed earlier purchases. 2025 Huxiu essay 2015 speech Yifeng prospectus Three Squirrels prospectus
• Limits, failure modes, and exception ledger — DCF can launder narrative into precision; venture terminal values dominate; a narrow competence circle can exclude discontinuities; and a low stated price can hide dilution or future capital need. Log sensitivity, failed forecasts, stale marks, missing securities, and decisions where price discipline correctly lost a later winner.
• Evolution and contradictions — This is the clearest public change: earlier claims that great companies overwhelm price sit in tension with the later margin-of-safety framework. Treat the later rule as a dated revision, not proof it was always used.
• Modern VC translation (as of 2026-08-23) — Transfer: price uncertain AI cash flows through scenarios inside a demonstrable competence boundary. Changed condition: model costs, platform bundling, regulation, and financing needs can shift faster than a terminal-value forecast. Signal: the thesis survives lower prices, more dilution, and slower adoption. Misuse: calling a high-growth company priceless. Hypothetical: decline an AI round whose ownership-adjusted base case needs both permanent model scarcity and no follow-on dilution. [researcher inference | strong]

Cross-model decision worksheet

These eleven models are an evidence-bounded reconstruction of Kathy Xu's public doctrine and documented behavior. They merge recurring synonyms rather than inflating every phrase into a separate model. A label such as [investor-stated | strong] describes attribution and support for the stated rule; it does not certify investment performance. The economic edge remains provisional because complete portfolio, pass, write-off, security, and cash-flow denominators are unavailable. investment philosophy, edge thesis investments, evidence boundary

Before an investment, answer all eleven questions in writing:

1. What does the full category map show, and which substitute is easiest to miss?
2. Which customer segment is least represented in the team's intuition, and what direct evidence changed the model?
3. What behavior proves founder learning and integrity, and what control would catch failure?
4. What capability does the proposed capital buy, by what date, and what ends support?
5. Who controls product/supply, distribution, and customer mindshare today and at scale?
6. What is the total-loss, correlation, reserve, and partner-attention effect on the fund?
7. Can the company survive 24 months without outside capital, and which follow-on gate is hardest?
8. At the next liquidity event, what facts would make selling, holding, or distributing correct?
9. What single recruiting, metric, or brand intervention has a measurable counterfactual?
10. What will the decision record allow the team to learn even if the outcome takes ten years?
11. What bear, base, and bull values follow from the actual security and ownership, and what maximum price preserves the required margin of safety?

Exception ledger and anti-hagiography controls

Compact glossary

• Forest: the complete category and substitute set considered before selecting a company.
• Tree error: becoming attached to one attractive company before comparing the category.
• Competence circle: the boundary within which the investor can recognize relevant facts and unknowns.
• Margin of safety: the gap between price and a conservative scenario value required to absorb uncertainty.
• Consumer immersion: structured observation used to update a market model.
• Three controls: product/supply, channel/distribution, and mindshare/brand.
• Evidence gate: a dated, measurable condition for releasing more capital or continuing support.
• Attention budget: the partner time a position can consume without degrading the rest of the portfolio.
• Financing winter: a period in which outside capital is scarce and survival must come from runway or insiders.
• Second decision: a fresh hold, sell, follow-on, or transfer underwriting.
• Intervention counterfactual: the likely company path without the investor's specific action.
• Decision journal: a contemporaneous rationale, forecast, dissent, and kill-criteria record.
• Exception ledger: adverse cases and conditions under which a model must not be applied.

Four operating checklists

Category and consumer

• Full company/substitute map dated before the favorite is chosen.
• Historical analogy and map-breaking difference both stated.
• Consumer sample spans relevant geography, income, language, and use case.
• Interview evidence reconciled with retention, payback, and contribution data.

Founder and governance

• Integrity and learning tested through independent behavioral references.
• Metrics reproduced from primary systems, not screenshots or narrative.
• Related parties, cap table, cash controls, legal history, and data rights reviewed.
• Board escalation, founder transition, and information rights prewritten.

Capital and portfolio

• Check, security, ownership, dilution, reserves, and cumulative exposure modeled.
• Each tranche purchases a dated capability or uncertainty reduction.
• Total-loss, correlation, currency, regulation, and partner-attention stress run.
• Hold/sell/distribute analysis refreshed at every liquidity event.
• Bear, base, and bull values use the actual security, ownership, dilution, and follow-on need.

Learning and evidence

• Original memo, probabilities, dissent, and kill criteria immutable.
• Every return claim has a cost-to-distribution bridge or is labeled not computable.
• Investor, firm, founder, co-investor, vehicle, regime, and luck attribution separated.
• Misses, ordinary cases, write-offs, passes, and exceptions reviewed with winners.

Suggested study path

1. Read the investment philosophy for the 17-part operating system and strongest skeptical case.
2. Use writings and talks to identify what Xu actually said, when, and through which access grade.
3. Read investments and adverse and miss postmortems to separate security facts, reported outcomes, misses, and attribution.
4. Apply Models 1–3 to a new opportunity without looking at its latest valuation.
5. Apply Models 4–9 to the same opportunity and write a check, reserve, governance, intervention, and exit plan.
6. Apply Model 10 one year later; score the forecast before rewriting the story.
7. Apply Model 11 to the bear, base, and bull cases; name the price that preserves the margin of safety.
8. Resolve locators and source-family limits in sources before citing a claim.

Evidence gaps and candidate models

• “7-2-1” category allocation. This broad portfolio slogan lacks a public base-rate or security-level denominator and is merged into Models 1, 4, and 5 rather than admitted separately. [researcher inference | strong]
• Roughly 2,000 consumer interviews. The count is self-reported and unaudited; it is an implementation claim under Model 2, not a distinct model. 2022 alumni interview [investor-stated | strong]
• “Three weapons.” Recruiting, metrics, and brand are already admitted together as Model 9; splitting them would inflate synonyms without independent decision rules. [researcher inference | strong]
• AI data-flywheel or blitz rule. Candidate only. The 2026 Apple episode has audio and chapter metadata but no repository transcript, the legacy XYZ Robotics PDF is unavailable, and public current-deal records do not expose security, ownership, governance, or outcome. 2026 Apple episode [investor-stated | provisional]
• “Price does not matter.” This is not admitted as a rule: later DCF and margin-of-safety language contradicts or supersedes it. Model 11 preserves the statement as an evolution counterexample. 2025 Huxiu essay [researcher inference | strong]

What the evidence supports

The strongest defensible conclusion is narrower than the public legend. Kathy Xu has a long, filing-visible career across Baring-era and Capital Today investments; her direct record repeatedly emphasizes category mapping, consumer observation, founder learning, concentration, financing support, and long holding. Selected filings establish meaningful ownership, governance, and partial sale facts. They do not establish a complete personal or fund return record, and the public corpus cannot show that the reconstructed method caused superior results across the whole portfolio. [researcher inference | strong]

The package therefore supports a decision system, not a greatness ranking. Its best evidence is the combination of direct statements with independent security, ownership, court, regulator, and issuer records. Its central economic claim remains provisional: consumer/category research plus founder assessment, concentrated ownership, gated follow-ons, and bounded operating help may create an edge in China consumer and internet investing, but audited comparative cash flows, passes, write-offs, and role allocation are missing. profile, edge thesis deal denominator

The reviewed package contains 49 annotated source routes, three writing-form records, 18 deduplicated appearances, 18 exact-schema notes, 44 included deal/miss paths, six no-hindsight case labs, 11 admitted mental models, and three completed peer comparisons. These are public-record counts, not a career or performance denominator.

What could falsify the thesis

The proposed edge would strengthen if contemporaneous files showed the same category map, consumer sample, founder tests, ownership logic, follow-on gates, dissent, and stop rules before both winners and ordinary or failed decisions. It would strengthen economically if a complete, vintage-adjusted cash ledger showed competitive net results after the largest true contributor was removed and after allocating decisions across Xu, partners, vehicles, and founders.

It weakens if the rules occur mainly in retrospective winner stories; if passes, ordinary holdings, and losses show no comparable process; if access, vintage, regulation, or founder execution explains the visible outcomes; if follow-ons were not gated; or if the team rather than Xu owned the decisive functions. It fails as a demonstrated economic edge if complete fund data show ordinary net returns, weak calibration, or outlier dependence. [researcher inference | strong]

Attribution and denominator audit

The legal-holder chain controls economic attribution. A GP-control filing may establish authority over a holder but not personal ownership of every asset or personal cash proceeds. A financing total belongs to the company or syndicate unless an allocation is disclosed. A board seat establishes a role at a date, not incremental causal value. A company outcome remains founder/team/company credit unless a dated investor action and counterfactual are documented. [researcher inference | strong]

The current public-position control is equally narrow. Capital Today Evergreen Fund's Q2 2026 13F reported $129,498,857 across PDD, Dingdong, and KANZHUN; PDD's $105,648,105 was 81.6% of that statutory U.S.-listed long-equity slice. It is not current AUM, the complete fund, cost, or realized return. SEC 13F cover information table

Transfer limits for modern VCs

• Category mapping transfers; a 2000s China consumer winner-take-most assumption does not. Modern maps must include open-source substitutes, platform bundling, data rights, security, and regulation. [researcher inference | strong]
• Direct observation transfers; a self-reported interview count does not validate a sample. Enterprise AI requires workflow shadowing, permission/error review, and behavior after novelty. [researcher inference | strong]
• Behavioral founder diligence transfers; personal admiration does not. Security disclosure, benchmark integrity, data provenance, and enforceable controls now belong in the test. [researcher inference | strong]
• Concentrated attention and reserves can transfer only inside a vehicle's loss, duration, liquidity, correlation, and partner-capacity limits. The officially reported evergreen tranche fund sizes and source-reported 28-year design cannot be assumed for a ten-year fund. [researcher inference | strong]
• Aggressive capitalization transfers only through prewritten unit-economic, governance, and abandonment gates. Cheap compute or fundraising momentum is not a moat. [researcher inference | strong]
• Long holding transfers as repeated re-underwriting, not identity. Price, security, dilution, opportunity cost, regulation, and continuation conflicts make every liquidity event a second decision. [researcher inference | strong]
• AI-specific doctrine remains provisional. The 2026 audio and chapter index are accessible, but no repository transcript supports a detailed current AI rule. 2026 Apple episode [investor-stated | provisional]

Unresolved evidence

• complete audited fund schedules, including every investment, pass, write-off, follow-on, transfer, sale, residual, fee, carry, and LP distribution;
• the JD.com bridge among the first reported $10 million, later reported $18 million, disclosed securities and sales, $2.6 billion reported value, and $2.9 billion reported cash-out figure;
• security and cash bridges for Ganji, Meituan, Xingsheng, Dingdong, BOSS, ChinaHR, NetEase, Wahaha, and other repeated economic claims;
• original IC memos, recorded discussions, dissent, pass reasons, re-entry triggers, postmortems, and founder/co-investor corroboration of interventions;
• Capital Today LPAs, allocation policy, reserve policy, valuation method, fee/carry terms, succession, and final disposition of the source-reported proposed vehicle change;
• final closing evidence for Dingdong's conditional $717 million corporate transaction and any later effect on Capital Today securities or cash;
• a stable transcript for the 2022 Capital Allocators audio, the 2026 AI/robotics audio, the 2018 WSJ appearance, and the full attributed book chapter; and
• final adjudication or a sourced response for the Inoherb allegation, plus any nonpublic Real Kungfu transaction/recovery ledger.

Central-claim audit

Reproducible ordinary-claim audit

Selection rule: in each of the seven pre-review root documents, select the first independently checkable factual proposition from each of the first two different substantive sections, skipping headings, methodological commands, thesis conclusions already in the central ledger, and local-navigation sentences. This yields exactly two claims per document and 14 total.

Deal and team cross-checks

Allegation, adjudication, and outcome states

Source, writing, talk, deal, and model reconciliation

• The 49-row source inventory reconciles to three writing-form records and 18 deduplicated appearances; mirrors and translations are explicitly non-independent.
• Two accessible priority writings have two w-* notes. Sixteen processed appearances have 16 t-* notes; two metadata-only appearances have no content note.
• All 18 notes use the exact metadata grammar and six required H2s. Guazi is full after revalidation; the two full-audio notes remain partial because no transcript was processed.
• The deal ledger contains 44 unique included rows and four separately named exclusions. Its 20 E1 / 24 E0 split matches every synthesis denominator reference.
• Six case labs preserve known/unknown/believed/disputed before outcome and ask invest/pass, kill, terms/ownership, and next-check questions first.
• Eleven ## Model sections each contain the 12 exact field labels and three nested, attribution-labeled examples. Price/margin of safety is distinct; slogans and untranscribed AI claims remain in the candidate ledger.
• The three peer links resolve to status-confirmed completed canon entries: 043-gary-rieschel, 030-jim-breyer, and 039-masayoshi-son.

Corrections made in controlled review

The review began from rejection and corrected every reproducible defect found:

1. reconciled the profile denominator from 18/44 and 26 unresolved to 20/44 = 45.5% and 24 unresolved;
2. normalized all 18 note metadata rows and the deal source-ledger grades to exact A/B/C and full/partial vocabularies;
3. repaired three nonexistent mistakes.md links, the Yonghui adverse locator, and current-holdings links;
4. replaced stale Q1 holdings with the official Q2 2026 accession, exact values, and denominator limits;
5. corrected BOSS from an erroneous $18 million to roughly 18 months of runway;
6. separated JD's reported $2.6bn value from Forbes's $2.9bn cash-out wording and preserved the 144.4x arithmetic conflict;
7. changed the official evergreen amounts from commitments/AUM language to tranche fund sizes totaling $2.849bn and separated the source-reported term from an unreviewed LPA;
8. added Dingdong's official conditional corporate-sale filing without treating $717m as a completed sale or investor proceeds;
9. preserved the Dazu/Dazhu childhood conflict, removed an unsupported birthplace, and downgraded ChinaHR's roughly 800x claim to provisional paper value;
10. revalidated Guazi as full access and reconciled the spoken corpus to 11 full, 3 partial, 2 untranscribed audio, and 2 metadata-only appearances;
11. corrected exact quotations in the 2015 Chaos, 2014 Wharton, and 2025 methodology notes and removed the misplaced 2018 cognition quote;
12. converted every model's examples to three nested labeled cases, added the distinct price/margin-of-safety model, expanded all dated translations, and added model-candidate exclusions.

Twenty-point acceptance matrix

Mechanical and rendered verification

• Files and source maps: 8/8 root files and 18/18 notes are substantive. The 49 source rows reconcile to 4 writing/book routes, 23 talk/interview routes, 11 firm/filing routes, and 11 context/adverse routes.
• Schema and labels: 18/18 notes pass the exact metadata regex and six-heading contract. The package contains 474 valid two-axis evidence labels, 0 invalid labels, and 0 non-A/B/C source-ledger grades.
• Models: 11/11 models contain each of the 12 required fields, producing 132/132 field instances and 33/33 nested labeled case bullets. Every model has at least 9 full citation URLs and at least 3 cited domains.
• Quotes: the extraction found 159 smart-double-quoted spans across roots and notes; the longest English span is 11 words and the longest Chinese span is 23 Han characters. The 18 note quote sections contain 23 bullets and 21 retained smart-quoted excerpts; their maximum is 10 English words. No excerpt exceeds 25 words.
• Links: 79/79 local Markdown file/fragment targets resolve under the repository’s installed GitHub slugger. Live reopening of the 31 external URLs in the central/ordinary ledgers returned 28 direct 2xx/3xx responses; the known legacy Capital Today route returned 404, the Guangzhou court and SAMR routes failed the automated fetch. Those three retain explicit archival/intermittent-access boundaries and alternate records where available.
• Build and render: npm test --prefix site passes 15/15 tests after a production Astro build. The rendered Kathy Xu page exposes 8/8 document tabs, resolves 73/73 document fragments, contains no .md href, and renders the final review verdict.
• Peers and hygiene: the live index and TODO records show Gary Rieschel, Jim Breyer, and Masayoshi Son complete. Placeholder, non-Markdown artifact, whitespace, first-H2, exact-grade, and invalid-label scans emit no failures.

Final verdict

Adversarial review passed. The package proves that Kathy Xu's public record contains a coherent, evolving set of investment rules and selected consequential filing-visible roles. It provides a reproducible public-source deal denominator, no-hindsight exercises, attribution controls, and a falsifiable method suitable for study.

It does not prove career alpha, a full portfolio hit rate, audited JD.com or ChinaHR multiples, current AUM, personal proceeds, sole decision credit, intervention causality, or that a China consumer-era method transfers automatically to 2026 AI. Those are explicit open tests, not implied conclusions.

Writings, statements, and books

• ★★★ | 巴菲特股东大会回来读段永平 | https://www.huxiu.com/article/4330139.html | 2025-05-09 | authored essay, full | Only complete attributable investment essay found; valuation, DCF, safety margin, concentration, and holding period.
• ★★ | 今日资本四点声明（徐新朋友圈全文转载） | https://www.thepaper.cn/newsDetail_forward_28037065 | 2024-07-10 statement / 2024-07-11 report | public statement, full quoted reproduction | Primary-market activity, LP calls, personnel adjustment, and office rebuttal; organizational record, not philosophy essay.
• ★ | 《增长的本质》chapter “今日资本：什么变了，什么没变” | https://wulolife.com/products/%E5%A2%9E%E9%95%BF%E7%9A%84%E6%9C%AC%E8%B4%A8-%E7%A7%A9%E9%98%B5%E5%BC%8F%E5%A2%9E%E9%95%BF-%E7%A0%B4%E8%A7%A3-51-%E4%B8%AA%E5%A2%9E%E9%95%BF%E9%9A%BE%E9%A2%98 | c.2020 | book-chapter metadata only | Possible writing, but full text and original provenance are unavailable; cannot analyze content.
• ★ | “一个身价150亿的女人…” compilation | https://www.10100.com/article/132590255 | 2023-07-08 | editorial compilation | Repackages known talks under a Xu byline; recycled lead, excluded from writing breadth.

Talks, interviews, podcasts, and video

• ★★★ | 南京大学2014毕业典礼致辞 | https://alumni.nju.edu.cn/4e/94/c6795a216724/pagem.htm | event 2014 / published 2017-08-31 | full official transcript | Biography, three choices, passion, focus, role models; contains a ChinaHR date error.
• ★★★ | 不第一，毋宁死 | https://www.tmtpost.com/1033818.html | 2014 | full edited speech | Category leadership, JD/NetEase, check size, logistics, and platform concentration.
• ★★ | Wharton Global Forum Beijing coverage | https://news.wharton.upenn.edu/wp-content/uploads/2012/06/Wharton-Global-Forum-Beijing_Coverage-Wrap-Up-Report_20140702.pdf | 2014 | official event summary PDF | Partial direct voice on category leadership; no full transcript.
• ★★★ | 创始人该把时间花在哪儿? | https://pe.pedaily.cn/201501/20150115376929.shtml | 2015-01-15 | full speech transcript | Annual retrospectives, misses, pricing, category speed, JD, founder time allocation.
• ★★ | HKEX Women Exchange luncheon with Kathy Xu | https://www.hkexgroup.com/-/media/HKExGroup/csr/wex/documents/2015-luncheon-with-kathy-xu.pdf?la=zh-HK | event 2015-04-15 / posted 2015-04-28 | official summary PDF | Transition-era thesis, first-mover quote, talent-management language; partial.
• ★★★ | 传统企业怎么做互联网+ | https://www.hundun.cn/articles/d27604e1ea8ad332.html | event 2015-08-08 / published 2015-08-09 | full edited transcript | Consumer/channel change, JD expansion, price wars, governance, options, no ratchets.
• ★★★ | 徐新：资本寒冬，活下来就是好样的 | https://www.36kr.com/p/1721195167745 | 2016-09-22 | edited event transcript/report | Runway, repeat purchase, CAC payback, fundraising timing.
• ★ | SuperReturn Asia 2016 panel | https://media.telecoms.com/superreturnasia-old/superreturn-asia-2016-final-agenda.pdf | 2016-09-12 | agenda metadata | Appearance verified; no recording or transcript located.
• ★★ | Guazi event speech | https://www.sohu.com/a/201608173_313170 | 2017 | edited speech, full | Product-led founders, Ganji/Guazi, hiring and marketing attribution; full body revalidated 2026-08-23.
• ★★★ | 36Kr two-hour interview | https://www.10100.com/article/28402593 | 2018-04-19 | full edited interview mirror | Most complete pre-2020 method interview; winner study, founder tests, misses, Yonghui, competence circle.
• ★ | 36Kr interview English translation, part 3 | https://kr-asia.com/be-inspired-by-kathy-xu-founder-and-president-of-capital-today-in-her-journey-of-investments-3-3 | 2018-06-11 | translation | Same underlying interview; translation aid, not independent evidence.
• ★ | Capital Today “WSJ Speech” | https://www.capitaltoday.com/video_6.html | May 2018 metadata | official legacy metadata, now 404/shell | Discussion with Phred Dvorak identified; no stable recording/transcript recovered.
• ★★★ | 徐新眼中的王兴 | https://m.yicai.com/news/100022154.html | 2018-09-05 | full exclusive Q&A | Meituan concentration, special approval, founder traits, high-frequency essential trillion-market screen.
• ★★ | Ray Dalio book-launch panel | https://finance.sina.com.cn/hy/hyjz/2020-08-14/doc-iivhuipn8714177.shtml | 2020-08-14 | partial video/report | Pinduoduo research inputs and user interviews; no complete stable transcript.
• ★★★ | Capital Allocators Ep.243 — Venture in China at Capital Today | https://www.capitalallocators.com/podcast/venture-in-china-at-capital-today/ | 2022-03-28 | 53-minute audio, transcript paywalled | Full audio accessible; publisher topic map; no public transcript used.
• ★ | Capital Today podcast mirror metadata | https://www.capitaltoday.com/video_1.html | legacy page says May 2018 | metadata | Same Capital Allocators episode but wrong date; diagnostic duplicate only.
• ★★★ | 南京大学校友会访谈 | https://www.10100.com/article/21949359 | interview 2022-05-15 / mirror 2022-05-25 | full edited interview | Two biggest mistakes, three-method process, 2,000 interviews, household immersion, return concentration self-report.
• ★ | 21st Century Business Herald copy of NJU interview | https://www.21jingji.com/article/20220524/herald/f34557f814c06676252df6c437f88bf0.html | 2022-05-24 | syndicated text | Same underlying interview; does not expand denominator.
• ★★★ | 28年风投女王从芒格那儿学到什么？ | https://finance.sina.com.cn/money/fund/jjzl/2023-09-02/doc-imzkhpuq5016180.shtml | event 2023-08-26 / published 2023-09-02 | 39-statement edited talk | VC vs value investing, concentration, home-run frequency/intensity, learning and temperament.
• ★ | Capital Today official mirror of Munger talk | https://capitaltoday.com/zh/insights/article/10 | 2023 | syndication | Same underlying Smart Investor/Munger Academy event.
• ★★★ | 我的投资方法论 | https://m.sohu.com/a/989877148_100246910 | event 2025-12-13 / published 2026-02-25 | full edited transcript mirror | Current three-method framework, ByteDance/PDD misses, household diligence, portfolio-capacity claims, AI/LLM use.
• ★ | Forbes China canonical methodology page | https://www.forbeschina.com/investment/%E6%8A%95%E8%B5%84/71141 | 2026-02-25 | canonical transcript page | TLS certificate validation failed; Sohu mirror used for full content.
• ★★★ | 高能量 Vol.227 — 具身新贵与投资女王 | https://podcasts.apple.com/us/podcast/vol-227-%E5%85%B7%E8%BA%AB%E6%96%B0%E8%B4%B5%E4%B8%8E%E6%8A%95%E8%B5%84%E5%A5%B3%E7%8E%8B-%E5%88%9B%E6%96%B0-%E7%83%AD%E6%BD%AE-%E6%B3%A1%E6%B2%AB%E5%92%8C%E5%91%A8%E6%9C%9F/id1635271946?i=1000778117610 | 2026-07-24 | 2:32:36 audio + full chapter index | Current AI/robotics, data flywheels, bubble/cycle, concentration, missed-deal and training-cost thinking; no transcript claim.

Firm records, filings, and deal data

• ★★★ | Capital Today official homepage | https://www.capitaltoday.com/ | current 2026 | official firm page | “DNA” and exact Fund I/II/evergreen vintages and sizes.
• ★★★ | KANZHUN Form 20-F | https://www.sec.gov/Archives/edgar/data/1842827/000119312522127960/d221166d20f.htm | 2022 | SEC filing | Career chronology and Xu's director role since April 2015.
• ★★★ | CTG Evergreen/KANZHUN ownership disclosure | https://www.sec.gov/Archives/edgar/data/1854545/000119312522135422/d287975d20f.htm | 2022 | SEC filing | CTG share count and Xin Xu's control of the GP.
• ★★★ | Capital Today Evergreen Q2 2026 13F cover | https://www.sec.gov/Archives/edgar/data/1625502/000119312526340257/primary_doc.xml | period 2026-06-30 | SEC filing | Three positions and $129,498,857 total disclosed U.S. long-equity value.
• ★★★ | Capital Today Evergreen Q2 2026 13F information table | https://www.sec.gov/Archives/edgar/data/1625502/000119312526340257/56524.xml | Q2 2026 | SEC filing | PDD $105,648,105 / 1,385,004 ADS, Dingdong $22,927,111 / 11,879,332 ADS, KANZHUN $923,641 / 71,767 ADS; limited statutory denominator.
• ★★★ | Dingdong definitive agreement 6-K, Exhibit 99.1 | https://www.sec.gov/Archives/edgar/data/1854545/000119312526038494/dingdong_definitive_agre.htm | 2026-02-05 | SEC filing | $717m corporate consideration for substantially all China operations; shareholder, SAMR, carve-out, tax, and other closing conditions; not Capital Today proceeds.
• ★★★ | JD.com IPO prospectus | https://www.sec.gov/Archives/edgar/data/1549802/000104746914005115/a2220275z424b4.htm | 2014-05-21 | SEC filing | IPO date, securities and ownership context; controls over erroneous media chronology.
• ★★★ | Xin Xu/Capital Today JD Schedule 13G | https://www.sec.gov/Archives/edgar/data/1549802/000090445415000085/s13g_013015-jdcom.htm | 2015 | SEC filing | Primary ownership/control route; automated SEC access later blocked.
• ★★ | CNINFO transaction disclosure | https://static.cninfo.com.cn/finalpage/2018-11-07/1205584628.PDF | 2018-11-07 | listed-company filing PDF | Entity map and bounded prior-five-year penalty/litigation representation.
• ★★ | Capital Today legacy JD PDF | https://www.capitaltoday.com/download/news29.pdf | legacy / now 404 | official-hosted media reprint | Indexed transaction dates and quotations; cites other press, so not independent primary evidence.
• ★★ | Capital Today legacy XYZ Robotics PDF | https://www.capitaltoday.com/download/news41.pdf | 2026 / now 404 | official-hosted press PDF | Indexed direct quote on robots, labor shortage, and technical-team selection; archive retrieval lead.

Press, profiles, skeptical records, and outcome checks

• ★★★ | Forbes — Rainmaker | https://www.forbes.com/global/2008/0407/050.html | 2008-03-28 | full reported profile | Earliest high-quality chronology and contemporaneous process description.

• ★★★ | Forbes — 2019 Midas profile | https://www.forbes.com/sites/jilliandonfro/2019/04/02/investor-kathy-xu-rockets-to-2019-midas-list-top-ten-as-power-of-chinese-startups-grows/ | 2019-04-02 | full direct interview/profile | Deal pace, China focus, reported JD proceeds; contains date/check conflation.

• ★ | Forbes profile page | https://www.forbes.com/profile/kathy-xu/ | 2022-04-12 update | biography/database page | Deal leads but a surfaced version contains a JD IPO-year error.

• ★★ | 徐新巨大的ego，以及旷日持久的真理实验 | https://www.chuangkem.com/archives/123654 | 2020-05-21 | full critical essay | Key-person, evergreen incentives, late-stage strategy, and failure leads; causal claims require corroboration.

• ★★ | Yonghui 2018 annual report mirror | https://money.finance.sina.com.cn/corp/view/vCB_AllBulletinDetail.php?id=5279438&stockid=601933 | 2019 | company filing mirror | Yunchuang losses and earnings context; counterevidence to 2018 optimism.

• ★★ | The Paper — Yonghui Life closures | https://www.thepaper.cn/newsDetail_forward_5381931 | 2019-12-30 | reported outcome | Store closures and reported H1 2018 loss; entity/perimeter reconciliation still required.

• ★ | “撤销一级团队” rumor article | https://finance.sina.com.cn/roll/2024-07-09/doc-inccppey0112768.shtml | 2024-07-09 | commentary/hearsay | Defines the allegation; not evidence of truth.

• ★★ | The Paper — Xu/Capital Today rebuttal | https://www.thepaper.cn/newsDetail_forward_28037065 | 2024-07-11 | reported full statement | Bounded response and current primary investments; does not disclose team count.

• ★ | SuperReturn Asia current speaker bio | https://aqmen365.com/superreturnasia/speakers/kathy-xu/ | 2026 | marketing biography | Current-title/deal lead; $6bn AUM conflicts with official fund page.

• ★ | HKVCA 2026 forum program | https://www.hkvca.com.hk/wp-content/uploads/2026/05/GCPES26_FP.pdf | 2026 | official agenda PDF | Future/very recent appearance lead; no transcript in the searched record.

• ★★ | Sina 2007 Kathy Xu profile | https://tech.sina.com.cn/i/2007-01-22/18491347591.shtml | 2007-01-22 | retrospective profile | Reported household-savings ChinaHR investment and roughly 800x paper-value claim; no security, sale, or cash ledger.

Coverage assessment

• Inventory: 49 rows across writings, talks/video, books, firm pages, filings, profiles, skeptical commentary, and outcome records. Duplicates and translations are explicitly marked and never counted as independent underlying works.
• Name/date variants searched: Kathy Xu, Cathy Xu, Xu Xin, Xin Xu, 徐新, 今日资本, Capital Today, plus interview/采访, 演讲, 对话, podcast, transcript/文字实录, 文章, 随笔, 股东信, 信件, filetype:pdf, 诉讼, 处罚, 监管, 争议, 批评, each with decade/year sweeps. Identity collisions require firm/career anchors.
• Direct voice: one full authored investment essay; one reproduced public organizational statement; one metadata-only book chapter; 18 deduplicated worthwhile talks/interviews (11 full/near-full text, three substantive partials, two full audios without public transcripts, two metadata-only appearances).
• Archive/access routes tried: current and legacy Capital Today paths, publisher pages, SEC/issuer archives, CNINFO, university/event PDFs, Apple Podcasts, Capital Allocators, mirrors, and Internet Archive CDX. Legacy Capital Today news_10.html, news29.pdf, news41.pdf, and video paths now 404 or serve the redesigned shell; CDX timed out. SEC later returned an automated-tool block. Forbes China failed local TLS validation; a fully attributed Sohu mirror was used. No paywall was bypassed.
• Deal denominator: Final forensics discovered 48 public paths, retained 44 in scope, sufficiently reconstructed 20, left 24 unresolved, and excluded four; the true career/fund population and cash-flow denominator remain unknown, so economic coverage and hit rate are not computable. deal reconciliation
• Historical context routes: 2008 Forbes for contemporaneous China PE regulation/market practice; issuer IPO filings; HKEX/Wharton event records; later listed-company outcomes. Retrospective speeches must not substitute for decision-date evidence.
• Skeptical routes: exact Chinese/English lawsuit, enforcement, penalty, controversy, and failure searches found no credible major Xu/Capital Today enforcement case. The 2018 CNINFO representation is bounded, not a universal clearance. Stronger public skepticism concerns key-person dependence, evergreen incentives, strategy drift, new-retail outcomes, and missing return denominators.
• Highest-value retrieval leads: transcribe the 2022 Capital Allocators and 2026 高能量 audios with the repository-approved OpenAI route; retrieve the 2018 WSJ recording; obtain the 《增长的本质》 chapter; recover legacy Capital Today pages from a functioning web archive; and build a filing-based JD cash-flow/ownership chronology.
• Completed-peer comparison: The final synthesis uses only three status-confirmed completed peers—Gary Rieschel, Jim Breyer, and Masayoshi Son—as dimensional controls for China context, networks, concentration, vehicles, governance, and aggressive capitalization; no return league table is inferred.

Source notes