SOFTWARE · MODELS · SILICON HARDWARE · HUMANITY 2026-11-02

10. Primary Sources, Method, and Plain Glossary¶

How Every Fact in This Book Was Verified¶

We checked the claims against 26 legal, technical and research sources as of 2 November 2026. Added research informs the discussion of cooperation and risk; it does not turn our assumptions into measured results. Contact information is what the cited records print and may change. Source verification is distinct from approving a transaction.

Original WPS bytes and SHA-256 hashes are preserved untouched alongside converted PDF reading copies. Where GAO's server blocked automated byte archiving (HTTP 403), we note that limitation transparently. Throughout the book, primary legal facts carry bracketed citations, while our game-theoretic matrix and future scenarios are presented openly as analytical models rather than legal advice for a specific transaction.

Untouched Official Document Copies¶

Plain-Language Glossary for Normal Humans¶

Model weights (parameters): The billions or trillions of tuned numerical values learned during training that shape a model's outputs. Weights alone do not include every part of a reproducible or safe system.
Checkpoint: A saved snapshot of weights during or at the end of training, used to resume training, distil smaller models, or fine-tune for specialist tasks.
FLOPs (floating-point operations): Individual math steps performed during training. \(10^{26}\) FLOPs (a 1 followed by 26 zeros) is the frontier compute threshold in both China's proposed 266504X and America's 4E091.
TPP & HBM (the hardware bottleneck): Total Processing Performance (TPP) measures an AI chip's raw compute throughput, while High-Bandwidth Memory (HBM / DRAM bandwidth) measures how fast data feeds into the chip (capped at TPP < 21,000 and 6,500 GB/s in the January 2026 US China-review rule 742.6(b)(10)).
Chip tape-out (芯片流片): Sending a finished processor blueprint from a chip-design team to a semiconductor foundry to manufacture physical silicon wafers, explicitly offered automatic-licensing facilitation in Announcement 48's Footnote 2.
API / SaaS (the bakery window): Letting remote users send prompts to your cloud server and receive answers back, without handing over underlying model weights, source code, or chip designs.
Nash equilibrium: A set of choices in which no player can improve its own payoff by changing only its own choice. An equilibrium need not be a stalemate, safe, fair or socially best. Mutual walls are an equilibrium only under the assumptions of our example. GAME
Software license vs. export license: An MIT or Apache 2.0 license is a private copyright permission from creators allowing anyone to reuse code or weights. A Technology Export License (技术出口许可证 or BIS license) is a sovereign government permit allowing controlled technology or hardware to cross a national border.
Automatic licensing (自动许可): The streamlined "apply-and-go" (即报即出) government permit channel proposed in Announcement 48's footnotes for open-sourcing restricted AI, self-owned overseas clouds, overseas chip tape-out, and humanitarian biomedicine.
ECCN, CCATS & SNAP-R: In the US system, an ECCN (3A090 for AI chips, 4E091 for weights) identifies a controlled item; SNAP-R (snapr.bis.gov) is the free electronic portal where exporters file license applications and CCATS classification requests.

Complete Register of 26 Verified Primary Sources¶

Table 4 Primary legal, regulatory, and technical sources verified 2 November 2026¶
Code	Official Source & Scope Verified	Date	Primary URL & SHA-256 Prefix
`CN48`	MOFCOM & MOST, Announcement No. 48 of 2026 (Sections I, II & III)	2026-10-16	mofcom.gov.cn (`a5261a74...`)
`CN-A1`	Announcement 48, Attachment 1: 40 Proposed Technologies & 4 Footnotes (WPS)	2026-10-16	Original WPS (`213158bb...`)
`CN-A2`	Announcement 48, Attachment 2: Pre-Export Registration Form (WPS)	2026-10-16	Original WPS (`d65818c2...`)
`CNQA`	MOFCOM Spokesperson Official Q&A on Announcement No. 48	2026-10-16	mofcom.gov.cn (`37fc685a...`)
`CNTIER`	PRC State Council, Regulations on Administration of Tech Import & Export (Arts. 30-39)	2020-11-29	xzfg.moj.gov.cn (`39f4b9bb...`)
`CNMEAS`	MOFCOM & MOST Order No. 2 of 2009, Measures for Prohibited/Restricted Tech Exports	2009-04-20	mofcom.gov.cn PDF (`d14e5230...`)
`CNPLAT`	MOFCOM Unified Business System Platform Notice & Support (`+86-10-6787-0108`)	2026-11-02	mofcom.gov.cn (`5bc9740f...`)
`US26`	US BIS Final Rule on Semiconductor License Review (`91 FR 1684`; `742.6(b)(10)`)	2026-01-15	govinfo.gov (`d3b27863...`)
`US25`	US BIS Framework for Artificial Intelligence Diffusion (`90 FR 4544`; `4E091`)	2025-01-15	govinfo.gov (`6e202458...`)
`USNE`	US BIS Press Release Directing Non-Enforcement of AI Diffusion Rule	2025-05-13	bis.gov (`14a05e6e...`)
`GAO26`	US GAO Decision `B-337935` on AI Diffusion Rule & Congressional Review Act	2026-05-12	gao.gov (HTTP 403; web-verified)
`EAR734`	15 CFR Part 734: Scope of the EAR, Published Open Source (`734.7`) & Contacts	2026-11-01	bis.gov/ear/734 (`eeb76281...`)
`EAR742`	15 CFR Part 742: Control Policy, CCL Based Controls (`742.6(b)(10)`)	2026-11-01	bis.gov/ear/742 (`3eccf859...`)
`EAR744`	15 CFR Part 744: End-User and End-Use Based Controls (`744.11`, `744.23`)	2026-11-01	bis.gov/ear/744 (`ab19c823...`)
`EAR774`	15 CFR Part 774: Commerce Control List (ECCNs `3A090` and `4E091`)	2026-11-01	bis.gov/ear/774 (`01d3beba...`)
`SNAPR`	US BIS SNAP-R Electronic Filing FAQ & Contact Guide (`snapr.bis.gov`)	2026-11-02	snapr.bis.gov PDF (`e0c3bdee...`)
`TREAS`	US Treasury Outbound Investment Security Program & Oct 2026 Proposed Rule	2026-10-22	home.treasury.gov (`75f52eee...`)
`DOJ`	US Department of Justice National Security Division: Data Security Program	2026-11-02	justice.gov/nsd (`1bf01adb...`)
`OSI`	Open Source Initiative: Open Source AI Definition 1.0	2024-10-28	opensource.org (`a5442179...`)
`R1`	DeepSeek-R1 Official Repository & MIT License Record	2025-01-20	github.com/deepseek-ai (`875b282c...`)
`QWEN`	Qwen3 Official Repository & Apache 2.0 License Record	2025-04-29	github.com/QwenLM (`2b7cdab9...`)
`GD`	GuideDog Documentation & Book Publishing Engine (`v1.0.0-beta.10`)	2026-11-02	github.com/insanai/guidedog (`6fe7704f...`)

Research and Historical References Added in This Review¶

NTIA: Dual-Use Foundation Models with Widely Available Model Weights (July 2024). A government assessment of benefits, risks and monitoring recommendations for then-current open models, not evidence about every future model.
OPENRISK: On the Societal Impact of Open Foundation Models. A research position paper on marginal risks and benefits. We distinguish its argument from a measured causal effect of export controls.
GAME: A Course in Networks and Markets, Chapter 1. Definitions of games, best responses and equilibrium. Our particular scores and repeated-game calculation are the book's own example.
JEFFERSON: Thomas Jefferson to Isaac McPherson, 13 August 1813. The original source of the candle quotation. The analogy does not itself establish an economic effect.

Credits and Reuse¶

Written by Vikrant Rathore, with assistance from Jennifer Yuan, Vicky Ni, Ronak Rathore and Kanak Rathore.

Original text, editorial work and new illustrations are licensed under CC BY-SA 4.0. Attribution and sharing adaptations under the same license are required. Official documents, quoted material, fonts, software and other third-party materials retain their own copyrights and licenses. See the repository's LICENSE and NOTICE files for the terms and exclusions.

All seven language editions are editorial translations. They have not received independent professional legal or native-language review. Original legal texts prevail. Examples and game scores must not be treated as official classifications or empirical forecasts.

Optional: Checking the Game Example¶

The opening story can be understood without these calculations. Here we make the extra assumptions explicit for readers who want to check them.

Why the Ending Is Stable¶

The ranking 20 > 17 > 5 > 3 gives each country a reason to restrict whatever the other chooses. With two walls, either country sharing alone would move its score from 5 to 3. Mutual sharing is Pareto-efficient in this small game: no other listed outcome improves one country's score without lowering the other's. That does not mean either country receives its largest individual score. GAME

When Future Cooperation Matters¶

For the next calculation, we additionally treat the invented scores as comparable payoffs in each round. The numerical threshold depends on that scale, not just on the order of preferences.

Countries also interact repeatedly. In this simplified game, suppose cooperation continues while both share, but a deviation earns 20 once and is followed by 5 forever. Let \(\delta\) represent how much a country values the next round relative to this one. Cooperation is at least as attractive as that deviation when:

\[\frac{17}{1-\delta} \ge 20 + \frac{5\delta}{1-\delta} \quad\Longleftrightarrow\quad \delta \ge \frac{1}{5}\]

This is a calculation inside the assumed game, not a diplomatic forecast. It assumes an indefinite relationship, observable choices and a credible response to deviation. Errors, hidden actions or a known final round can undermine that strategy. The useful policy question is how to make continued cooperation credible, rather than treating mistrust as inevitable. GAME

Optional: Independent Review Is Not a Safety Guarantee¶

Open access can make independent testing and repair easier. A simple probability example explains one possible benefit. If a flaw exists with probability \(p_0\) and each of \(n\) teams independently detects it with probability \(q\), then:

\[P(\text{flaw remains undetected}) = p_0(1-q)^n\]

For a flaw known to exist and an assumed detection chance of one half, one team misses it with probability 50%, four independent teams with 6.25%, and ten with about 0.098%. These are conditional calculations. Real teams can share blind spots, lack testing resources or find a flaw without fixing it. An undetected-flaw probability is not a measure of total cyber, biological or societal risk.

Open weights can also help harmful actors modify a system or remove safeguards. NTIA's 2024 report considered both benefits and risks, recommending monitoring rather than restricting then-current open weights. Research on open foundation models stresses the need to compare the extra risk with existing alternatives. Neither source establishes a universal safety multiplier for openness. NTIA OPENRISK