The Huang Filecase file · v0.1

Method

Three layers

  1. Facts (F-) are dated events or documents. A fact is not an explanation. A news report counts as a fact only in the sense that it was published. What the report says goes in as claims.
  2. Claims (C-) record who asserted something, when, and through which source. When the claimant is an unnamed official, the claim says so (claimant_type: unnamed, and the body says "Unnamed-source claim").
  3. Theories (T-) are competing causal stories. Each one lists the facts and claims it explains, those it strains, those it ignores, and what would disprove it. Every theory has status open. The file does not rank them.

Fact status

status meaning
verified Primary document retrieved (official statement, filing, original post), or two or more independent outlets agree on the event itself
corroborated Reported directly by at least one established outlet, or by an official party to the event, without contradiction, but the primary was not retrieved or independently confirmed
contested Credible sources disagree
unverified Asserted only; nothing independent was retrieved

Claim status

unconfirmed · corroborated (others confirm independently) · partly-supported (some elements check out, others don't or weren't checked) · disputed (a credible party denies or contradicts it) · contradicted (later evidence shows it was wrong) · qualified (a later statement by the same party narrows it) · walked-back · opinion (a value judgement or political statement)

Rules

  • Quotes are exact and carry quote_source. If the original wording wasn't retrieved, the claim is paraphrased and isn't shown as a quote.
  • "X reported that Y said Z" is never written as "Z happened."
  • Widely repeated is not the same as proven. Unnamed-source claims stay labeled that way even when many outlets repeat them.
  • Every source note records how much was actually retrieved: full, snippet (search excerpt only), or paywalled.
  • Nothing is invented. If something can't be checked, it stays unverified and is listed in GAPS.md.
  • Dates are publication dates in US Eastern time unless a UTC time is given.

Topics added 2026-11-13

  • blackwell: the chip-forensics angle (DePue report, official allegations of Blackwell use, rebuttals).
  • export-controls: rules and enforcement history.
  • context: background and tangent material (Huang's biography, export-control history, earlier controversies, Nvidia–China business). It is filterable on the timeline and is not evidence for any theory unless a theory's mapping cites it.

Tangential ("random") items (added 2026-11-13)

Tangential items are in scope. These are stories that ride on the main event without being part of it, such as a lawmaker's household stock trades or tabloid side angles. They carry the topic tangent so the timeline can filter them in or out. They also carry subject topics such as markets, congress or media, the last for tabloid and press framing. The three-layer rules still apply: a tabloid's framing stays a claim unless independent records back it, and the record itself becomes a fact.

Known gaps (as of 2026-11-12)

  • Paywalled originals not retrieved: NYT English originals (Oct. 29, Nov. 6, Nov. 8, Nov. 11), FT (Nov. 6), WSJ (Oct. 30; board-ouster report Nov. 11), Bloomberg (Oct. 29 interview). Their content comes from NYT Chinese-edition translations (free), wire and aggregator summaries, and other outlets quoting them.
  • Search-snippet-only sources: several CNBC, Guardian, POLITICO and Anthropic pages (marked retrieval: snippet).
  • No named US official has confirmed the intelligence assessment on the record in any source retrieved here.
  • White House response to the NYT's "briefed and took no action" claim is known only from headlines (Mediaite, Washington Post). Not retrieved.
  • Second Nvidia employee: the Guardian refers to affairs with "two company employees". Only Kimberly Powell is named in retrieved sources.
  • Seed sub-claims not checked: the Jack Clark–Ana Swanson sourcing pattern, the Aspen report, Karnofsky's Anthropic role, the FT Anthropic exclusive (Sept. 2025), Amodei's Feb. Senate Banking meeting, Amodei's $1M to Public First Action, Public First ads praising Mast, most of the Anthropic hires list, whether Rice/Kinzinger/Magaziner/Kristof quoted McGuire's post, and whether it was the most-viewed post of the episode.
  • AnthroPAC dates: the tracker shows amounts but not dates. FEC filings not opened.
  • Alice Huang has made no public statement in sources retrieved (NYT says she did not respond).
  • Shareholder class action complaint not retrieved.
  • sources/raw/seed-x-2120746927100715509.raw.txt is the raw tool output (with tool wrapper lines). The clean copy is the .json.

Added 2026-11-13 (Blackwell forensics, side theories, context)

  • DePue forensics not independently replicated. The validators he cites (OpenAI, Anthropic, others) are unnamed. Neither the GitHub repo (github.com/macrogrants/checkpoint-forensics) nor the cited GitHub evidence (DeepGEMM PR 112, Ant cuLA GB300 logs) was opened.
  • DeepSeek-V4.1-Flash is unresolved. At 05:48 UTC Nov. 13 DePue posted that it "might be NVIDIA". Watch for a follow-up.
  • No responses found from DeepSeek, Moonshot, MiniMax, Ant Group, Nvidia, Commerce or China's MFA to the DePue report.
  • Paywalled or not read in the original: The Information (Nov. 12 DeepSeek/Litong story and the earlier Litong 240k-chip story), WSJ Nov. 12 (known only via emptywheel quotes), Culper Research's report, HK01, the C4ADS "Covert Compute" report, Kratsios's original X post, the SDNY indictment, and the BG2 podcast.
  • Legal question open: whether remote or overseas use of Blackwell by Chinese labs violates current US rules was not researched.
  • Side theories considered but not added: "smart money / short-seller hit job" and "psyop to push Jensen out". These circulate only from low-follower or anonymous X accounts with no evidence (for example, search results on Nov. 9–13), so they did not meet a credibility bar. Culper's thesis is captured as T-0010 instead.
  • Context gaps: Nvidia's China revenue share by year, the AI OVERWATCH / GAIN AI / Chip Security Act texts, Huang's April 2025 $1M dinner and $25M ballroom donation (asserted by emptywheel and Transformer, not independently sourced here), the Groq deal, and earlier NYT (Nov. 2025) and Bloomberg (Dec. 2025) Megaspeed reports.

Added 2026-11-13 (NY Post / Khanna trade)

  • The House periodic transaction report (PTR) for the Oct. 7 sale was not opened. Trackers disagree on the filing date (Nov. 2, Nov. 5 or Nov. 9) and on the September trade dates (Sep. 4 vs. Sep. 23, and sale vs. a dependent's buy). See F-0068 and F-0069.
  • Nvidia's all-time-high date (NY Post: one day before Oct. 7) was not independently checked.
  • Khanna's committee title ("top Democrat" on the China select committee, per the NY Post) was not verified.
  • The NY Post quotes the spokesperson as texting the "California Post". The original California Post item was not found.
  • Washington Sun (net worth) and Unusual Whales (most frequent trader) figures appear only as cited by the NY Post. The originals were not opened.
  • National File (S-0085) is a partisan outlet and only a search excerpt was read. It is used only for its own caveat and quotes.