Draft Audit

File audited: ../ai_views_blog.qmd

This audit checks the first draft against flow_test.md, style_contract.md, and acceptance_checklist.md. The draft is structurally sound, but it is still too compressed in places. The next revision should add concrete texture and smoother paragraph movement, not simply make every section longer.

Findings

1. The opening states the thesis clearly but starts too abruptly.

Lines: ../ai_views_blog.qmd:11-18

The opening has the right object: different mathematical tasks have different feedback structures. It also reaches feedback cost quickly. The weakness is that it reads like a sequence of scaffold claims. The first paragraph would be stronger if it began from a concrete contrast, for example a formal proof checker versus a conjecture or proof explanation, and only then named the broad question as too coarse.

Revision direction: Keep the claim, but let the contrast carry the paragraph. Avoid adding a throat-clearing sentence about AI discourse.

2. The “Cheap Feedback” paragraph is accurate but underdeveloped as the core mechanism.

Lines: ../ai_views_blog.qmd:20-26

This is the central mechanism of the essay. Right now it is only seven short sentences. It should become one more developed paragraph explaining why cheap failure changes search. The current sentence “It changes the economics of search” is correct, but it arrives too quickly.

Revision direction: Add one concrete pass through the loop: candidate generation, cheap rejection, retained successes, improved search. Use code/proof/simulation as examples, but avoid turning the paragraph into a list.

3. The proof section has the right tripartite distinction but still sounds slightly enumerative.

Lines: ../ai_views_blog.qmd:30-38

The generation / verification / digestion distinction is the right move. The three one-sentence definitions are too staccato. The unit-distance example helps, but the paragraph should better explain what digestion does operationally: it produces a proof that can be checked by humans, taught, modified, and recognized elsewhere.

Revision direction: Merge the three definitions into a smoother sentence or two. Add one phrase explaining why the human-written account mattered beyond shortening.

4. “Importance Oracle” is one of the strongest sections conceptually, but needs one concrete example of useless truth.

Lines: ../ai_views_blog.qmd:42-52

The section makes the right distinction between correctness and importance. It would benefit from a concrete example of the space of formally true but uninteresting statements. Without that, “Most true statements are not worth much attention” is correct but abstract.

Revision direction: Add a small example: arbitrary variants, technical lemmas with no reuse, or statements produced by changing constants/hypotheses without opening a method. Keep it short.

5. “Synthesis Debt” is clear, but the term needs to earn its name.

Lines: ../ai_views_blog.qmd:56-64

The paragraph defines the term and uses it correctly. It should give one more sentence explaining why “debt” is the right metaphor: the cost is not paid when the result is proved; it is paid later when people need to teach, reuse, or reorganize it.

Revision direction: Add that cost-transfer sentence. Do not dramatize it as a crisis.

6. The community and education sections are good but may be too separate.

Lines: ../ai_views_blog.qmd:66-84

These sections have a natural continuity: communities preserve working knowledge, and training is how new people enter those communities. Separated headings may make the post feel more segmented than necessary.

Revision direction: Consider merging them under a heading like Communities And Training, or keep two headings but add a stronger handoff sentence at the end of the community section. If merged, use two substantial paragraphs inside one section.

7. “Measurability Distortion” contains scaffold residue.

Lines: ../ai_views_blog.qmd:88-96

The section is right, but two sentences sound like planning notes:

  • “This does not require a broad attack on benchmarks or institutions.”
  • “The mechanism is enough.”

These are useful constraints for the writer, but they should not appear in the final prose.

Revision direction: Replace them with object-level prose. For example: benchmarks can remain useful while still changing incentives when they become the easiest visible measure of progress.

8. The conjecture section is promising but too compressed around Erdos.

Lines: ../ai_views_blog.qmd:100-107

The Erdos point is important and probably deserves one more sentence. Right now the paragraph asserts the example but does not quite explain it.

Revision direction: Say that if machines close many supplied conjectures, the scarce human act becomes selecting conjectures that were worth supplying. Avoid heroic language about question-makers.

9. The science map currently risks becoming a list.

Lines: ../ai_views_blog.qmd:111-120

This section is necessary, but the current paragraph lists fields one after another. It needs a stronger organizing sentence that says the field name is not the unit of analysis; the feedback loop inside the field is.

Revision direction: Keep the examples but compress them into fewer, more relational sentences. The biology contrast is probably enough, followed by one mathematical contrast and one slow-feedback example.

10. The ending is structurally right but too compressed for the final emotional landing.

Lines: ../ai_views_blog.qmd:124-131

The ending returns to the feedback-loop test, which is correct. It currently packs plateau, AGI, labor, and final thesis into one paragraph. That is too much for the final section.

Revision direction: Split the final section into two substantial paragraphs: one for plateau/AGI, one for labor and the final return. This is one of the few places where breaking the one-paragraph-per-section rule is justified.

Section-Level Decisions

Recommended merges:

  • Consider merging The Community Layer and Acquisition Asymmetry. They are one argument: working knowledge lives in communities, and training is how that knowledge is reproduced.

Keep separate:

  • The Importance Oracle and The Conjecture Economy. They are related but do different work. The first explains missing evaluation signal; the second explains institutional value shift.

Possibly rename:

  • The Importance Oracle may be too jargon-like for a blog heading. Options: What A Proof Checker Cannot Check, Correct Is Not The Same As Worth Proving.

  • Measurability Distortion may also be too abstract. Options: When The Scoreboard Becomes The Target, What Gets Measured Gets Trained.

Tone Audit

The draft avoids the worst failure modes:

  • no hype language;
  • no manifesto voice;
  • no grand claims about civilization;
  • no claim that AI cannot do serious mathematics;
  • no claim that all student AI use is cheating;
  • no unsupported timelines.

Remaining tone risks:

  • Some paragraphs still read like expanded scaffold cards.
  • Several sentences are too short in sequence.
  • A few writer-facing guardrails remain in public prose, especially in the measurability section.
  • The draft sometimes names the mechanism before making the reader feel the example.

Evidence Audit

Sources are used conservatively. The draft cites only verified entries in ../ref.bib.

Citation status:

  • @Avigad2026: used appropriately for AI-assisted mathematics and mathematical understanding.
  • @Alon2026: used appropriately for the unit-distance example.
  • @Schwer2026: used appropriately for mathematical practice/community meaning, but the draft should avoid leaning too much on it without rereading exact claims.
  • @Commelin2026: used appropriately for community, education, and institutional concerns.
  • @Klowden2026: used appropriately for education/social response, but exact claims should be checked before making stronger claims.
  • @Gwern2025: used only at the end for assistant-to-replacement pressure, which is appropriate.

Weak evidence areas:

  • Feedback-cost map beyond mathematics is mostly internal synthesis. Keep it as the author’s model unless Max Welling or another source is verified.

  • Conjecture economy currently rests on Avigad/Alon plus internal synthesis. Do not cite MathOverflow until verified.

  • Measurability distortion should not cite the Leiden Declaration until verified.

Revision Plan

  1. Smooth the opening and cheap-feedback paragraphs so the examples carry the mechanism.
  2. Expand proof/digestion and importance with one concrete example each.
  3. Replace scaffold-residue sentences in measurability.
  4. Decide whether to merge community and acquisition asymmetry.
  5. Strengthen the Erdos paragraph by one or two sentences.
  6. Compress the science map into fewer examples with clearer feedback contrast.
  7. Split the final section into plateau/AGI and labor/ending.
  8. Re-run:
    • citation validation;
    • banned-tone scan;
    • Quarto render.

Target after revision:

  • 2,200 to 3,000 words.
  • Same argument.
  • More concrete examples.
  • Less scaffold cadence.