method

the lanes

the corpus was produced by two parallel agent research lanes — history-and-people (the fund, the fleet, the founders, the papers) and coverage-and-idea (the shock, the scrutiny, the debates, the whale-funds-the-lab model) — each writing a lane report against one shared scratch source catalog before synthesis. the pattern is the company-research formula proven on the portfolio's earlier dossiers: lane-prefixed source ids frozen upfront, per-claim confidence, and a declared deduplication owner. 245 lane-minted ids merged into 230 canonical sources.

the evidence rules

1. every event, quote, and claim carries at least one source id.
2. dates keep their precision — a year, a month, or a day — and approximate dates are labeled rather than sharpened.
3. confidence stays honest: confirmed, reported, and inferred are distinct claims, not stages of the same claim.
4. sources are typed by role — primary, interview, reporting, archive, community — so a claim's foundation is visible before it is read.
5. sentiment labels describe the coverage, not the truth of it.

known gaps

1. the true compute inventory is unknown — ~10,000 a100s are documented and 2,048 h800s are self-reported; semianalysis's ~50,000-gpu estimate and the smuggling allegations remain analyst and prosecutor claims, not registry facts.
2. the lab's total spend is undisclosed — r&d budget, cumulative training cost, and the fund's internal transfer pricing between trading and research are all private.
3. r2's status is unresolved — reported stalled, then reported failed on ascend hardware; it was never announced, so there is no official cancellation to cite.
4. the 2026 round's terms conflict across reporting — ~rmb 300b at open, ~$52b at close per forbes, higher and lower figures elsewhere; no filing exists.
5. chip sourcing and export-control claims stay attributed — nvidia says its chips were lawfully acquired; the state-department military-linkage allegation is a named-official claim, not a finding.
6. the distillation question is open — openai's evidence was never published, anthropic's 2026 disclosure is one-sided, and deepseek's denial sits in the nature referee exchange.
7. censorship claims layer differently — app-serving refusals, in-weights tendencies, and hosted-serving behavior are separate findings, kept separate here.
8. headcounts, departures, and the ~130/~200 staff split come from anonymous-sourced reporting; the talent-war names are reported, not rostered.
9. the purported investor-call transcript (the "64 quotes" leak) is unverified — cataloged as a leak, not a transcript.
10. liang's pre-2015 trading record is self-reported — the rmb 80,000 origin story has no independent verification.

corrections

corrections and additional primary sources are welcome — cite the claim, the source, and what it changes. the catalog is designed to be audited, not just read.