References for the AI exposition project
What to read, in what order, and what has actually been checked
One source checked, two transcribed second-hand, the rest unread. The Tao passages quoted in the specification were read against the PDF on 2026-09-15, so entry 1 is ✅. The Thurston sentence and the four Leiden recommendations are quoted as Tao quotes them and have not been checked against their sources, so entries 2 and 9 are ⚠️. The two modeling-report guidances, entries 6 and 7, are named from memory of their titles and have not been opened; the question posed against them in Section 7 of the specification is unanswered.
The working document is the specification, and the folder’s other documents are on the project index.
Status markers
- ❌ Not checked. No claim in the working document has been verified against this source. Where the working document characterizes it, the characterization comes from memory or from another source’s citation.
- ⚠️ Transcribed, unverified. A quotation or claim in the working document came from this source at second hand and has not been checked against the source itself.
- ✅ Checked. Verified against the source.
On a reading-queue entry, ✅ means the question posed against that entry has been answered, by reading the source or by deciding it is out of scope. The note on the entry says which.
On a source entry below the queue, ✅ keeps its stricter meaning: the claim the specification draws from it has been checked against the source. Change a marker when the check is done, not when the paper is opened.
How to read these
Read against the question under each entry rather than front to back. The general method is on the Reading Papers page. Tao’s essay is twelve pages and the passages that bear on this project are in one section, so the reading is short; the two regulatory guidances are longer and the question against each is a lookup.
Notes go one line per question, carrying an answer and the page or section it came from, plus anything that contradicts the specification.
Reading queue
Ordered so that each makes the next easier to read.
1. Tao 2026 — Mathematics in the age of AI ✅
arXiv, submitted 2026-08-17. https://arxiv.org/abs/2608.16753
Question. What does the essay say AI-drafted exposition gets wrong, and does it say what to do about it?
Answer, 2026-09-15. Section 6, pages 6 to 7. Four failures, quoted in Section 3 of the specification: misplaced emphasis, no prior literature, no overview, and uniform friction. Footnote 4 on the flawless formatting reads “One can argue that they are too flawless.” The sentence after the Thurston quotation says authors assist digestion “by describing the insights, the false starts, and the stories from the period when they were working on the problem”. Section 8 carries recommendations to the community; the one addressed to authors asks them to disclose tool use and give “precise and complete references to previous results”, which is failure 2 as an instruction. Goal 6.5 and Figure 4, the five-stage pipeline with digestion as a dotted arrow, are on page 8, and the caption says only the first stage was ever an explicit goal of the community. Nothing in the essay proposes an instruction for failures 1 or 4. Section 7, page 9, lists four stages at which AI-generated proofs will pile up and names the transition from proof scarcity to proof abundance; its translation is in Section 3 of the specification. Section 8, pages 10 to 11, quotes four Leiden Declaration recommendations and closes with the rule of thumb that a result whose authors cannot give a clear expert-level talk on it should not be published; both are in Section 3 of the specification. Section 9 and Appendix A were not read.
2. Thurston 1994 — On proof and progress in mathematics ⚠️
Bulletin of the American Mathematical Society 30(2).
Question. Beyond the sentence Tao quotes, what does Thurston say about how mathematical understanding is transmitted, and does any of it translate to a modeling report?
Status. Not opened. The quotation in Section 1 of the specification is transcribed from Tao’s page 7 and matches his text; whether it matches Thurston’s is unchecked.
3. The writing guide in this repository ✅
.github/skills/writing-for-andy/SKILL.md in this repository.
Question. Which of the four failures does it already have a rule or observation for?
Answer, 2026-09-15. Read in full this session. The mapping is the last column of the table in Section 4 of the specification: O8 and O14 for failure 1, O3 for failure 2, O9 and rule 10 for failure 3, O16 and rule 10 in part for failure 4. Its Part 5 proposes recording what was tried and rejected, which is instruction 5 in the specification.
4. Passo 2024 — What’s wrong with AI-generated docs ❌
https://passo.uno/whats-wrong-ai-generated-docs/
Question. The writing guide cites this for the catalogue of LLM pathologies (hedging, verbosity, over-elaboration). Does the catalogue include the emphasis failures, or only the surface ones? If it stops at the surface, the guide’s tic list has already absorbed everything it offers.
5. Diátaxis ❌
Question. The four documentation kinds in the writing guide’s contract come from here. Does the framework say anything about where difficulty should be visible within a document, or only about which kind a document is? Section 5 of the specification proposes resolving the friction tension by tier, and this is where that proposal would find or fail to find a precedent.
6. EMA 2007 — Guideline on reporting the results of population pharmacokinetic analyses ❌
European Medicines Agency, Committee for Medicinal Products for Human Use.
Question. Does the required report content ask for the decisive modeling choice to be named, or for comparison with published models of the same drug? Open question 3 in the specification.
7. FDA 2022 — Population pharmacokinetics, guidance for industry ❌
US Food and Drug Administration.
Question. Same as entry 6. The two together are the published standard for what a modeling report must contain, and if either already asks for what instructions 1 to 3 propose, the instruction has a precedent outside this repository.
8. Tao 2021 — Exploring the toolkit of Jean Bourgain ❌
Bulletin of the American Mathematical Society 58.
Question. Tao’s Figure 3 is an annotated Bourgain page, offered as the case where friction taught the reader. Does this essay say what in Bourgain’s writing carried the signal? Lowest priority; it bears on failure 4 only and the specification does not cite it.
9. Leiden Declaration on Artificial Intelligence and Mathematics ❌
June 2026, endorsed by the International Mathematical Union. Cited by Tao as his reference 11.
Question. Tao quotes four of its twenty-three recommendations, the ones addressed to individual mathematicians. Do the others, addressed to organizations, funders and policymakers, say anything about exposition or review that translates to how a modeling group runs its reviews?
Sources
Every source the specification cites, with the claim drawn from it and a marker. Numbers match the queue above.
- 1. Tao 2026. ✅ The four failures, the footnote, the digestion sentence, Goal 6.5, Figure 4, the four abundance stages, the four recommendations and the rule of thumb, quoted, redrawn or paraphrased in Sections 3 and 5 of the specification. Checked against the PDF on 2026-09-15.
- 2. Thurston 1994. ⚠️ The measure-of-success sentence in Section 1 of the specification. Transcribed from Tao’s quotation.
- 3. The writing guide. ✅ Every O-number and rule number in the specification. Read in full on 2026-09-15.
- 6. EMA 2007. ❌ Named in open question 3 of the specification as the published report standard. Not opened.
- 7. FDA 2022. ❌ Same as entry 6.
- 9. Leiden Declaration 2026. ⚠️ The four recommendations in Section 3 of the specification, transcribed from Tao’s quotation of them. Not opened.