Add callouts for Fine and Technical Points mentioned without one

[d-morrison]

Requested by Ezra · project thread

Before: chapters 4, 15, 16, 21 and 23 mentioned 11 of their own Fine and Technical Points without the .callout-note that .github/copilot-instructions.md requires.

After: each of those boxes has a callout. Each one is paraphrased and titled with the book's numbering and box title (19 Aug 2026 revision). Re-running the issue's gap loop on current main gave this list. Since the issue was filed, TP 23.3 got a callout, so it is no longer a gap. Chapter 5's "Fine Point 5.3" shows up only because the loop misses its single-quoted title; that callout already exists.

Chapter	Box	How
4	TP 4.3: Relation between marginal and conditional causal risk ratios	wraps @prp-marginal-rr-condition and its proof, with a short intro
15	TP 15.1: Balancing Scores and Prognostic Scores	new callout. Takes over the speaker-notes block that held this material
16	TP 16.5: Proof of the General Homogeneity Condition	wraps the existing proof, with an intro
16	TP 16.7: Monotonicity and the Effect in the Compliers	wraps the existing proof. Brings in the attribution and surrogate-instrument notes and adds the book's corollary
21	TP 21.5: A Plug-In K+2 Robust Estimator	new: explains why the extra covariates make the correction terms vanish (score equations)
21	FP 21.3: G-Estimation with a Saturated Structural Nested Model	wraps @exm-g-estimation-saturated, with an intro
21	TP 21.8: A Closed Form Estimator for Linear Structural Nested Mean Models	new: derives the closed form and its multiply robust version. A notes-level fact-check covers the book's display
21	TP 21.11: A Big G-Formula Proof of the Front Door Formula	new callout after @exm-front-door-big-g, adds the coupling argument
21	TP 21.12: A Front Door Formula Proof Using d-Separation of Treatment Nodes on SWIGs	new: SWIG property and proof. Replaces the speaker-notes remark
23	TP 23.1: Proof of the mediation formula	new callout after the proof
23	TP 23.2: When the mediation formula is the g-formula	new callout. Takes over the speaker-notes block

How:

• Where a wide derivation would overflow inside a nested callout, the callout sits beside the derivation and links to it instead of wrapping it. TP 21.11 and TP 23.x are handled this way.
• I shortened a few align annotations, and in two places moved them into prose, so the nested proofs fit.
• Fact-check (TP 21.8): the book's closed-form display puts $A_{i,k}$ outside the sum and leaves $A_{i,j}$ out of $S_{i,k}$. Substituting the linear blip into $H_k(\beta) = Y - \sum_{j \ge k} A_j \gamma_j$ gives $S_k = \sum_{j \ge k} A_j R_j$. The two forms agree only when $K = 0$, so the callout uses the derived form and a notes-level warning explains the difference.
• Copyright check: I compared 8-grams of all added text against pdftotext of the book. The only shared runs are the box titles and one run of math symbols.
• Each changed chapter was rendered to HTML on its own, and I checked screenshots of every new callout.
• The semantic-line-break check passes. Spelling shows nothing new compared with main.

Closes #93

🤖 Generated with Claude Code

https://claude.ai/code/session_01Gh7L9vJhwrpvpBhzYMdjPK

---

Generated by Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[github-actions]

Claude finished review — View run

---

All content is uniquely placed, no duplication. The fix is clean and complete.

Review

This round (commit 49d39be, merged into be79a28) un-nests the four Technical/Fine Point callouts added in the prior round (TP 4.3, TP 16.5, TP 16.7, FP 21.3) from ::::-wrapping-::: nested divs into sibling ::: divs, per the commit message's stated goal of fixing DOCX rendering.

Verification performed:

• Confirmed the repo's _quarto-website.yml defines a real docx: format target (_quarto-website.yml:244) using the coatless-quarto/custom-callout and sun123zxy/callouty-theorem Pandoc filters — a plausible and real reason nested nested-div callouts could fail to render correctly to DOCX, since filter-based callout rewriting is known to be fragile with nested Pandoc fenced divs.
• Checked every changed file (chapters/04-effect-modification.qmd, chapters/16-instrumental-variable-estimation.qmd, chapters/21-g-methods-time-varying.qmd) for leftover :::: fences — none remain (grep -n "^::::" chapters/*.qmd is empty), so the un-nesting was applied completely and consistently to all four blocks touched by the prior round.
• Traced each un-nesting hunk by hand: in all four cases the outer ::::/inner ::: pair becomes two sibling :::/::: pairs with fences correctly balanced (opening/closing counts match); no stray or mismatched fence was introduced.
• In TP 16.7, a paragraph ("The proof assumes a causal instrument…often doubtful.") that previously trailed after the inner .proof div (inside the outer callout) was moved to before the proof div, now inside the callout box itself, with a one-word wording tweak ("The proof assumes" → "The proof below assumes"). Confirmed via grep that this paragraph's distinctive closing sentence appears exactly once in the file — the content was relocated, not duplicated or dropped.
• Each of the four un-nested callouts gained one new bridging sentence ("@prp-marginal-rr-condition below works this out…", "The proof below fills in each step.", "@exm-g-estimation-saturated below carries this out for @tbl-seq-rand.", plus the TP 16.7 reorder) to preserve the narrative link to the sibling block that immediately follows. Confirmed each referenced anchor (@prp-marginal-rr-condition, @exm-g-estimation-saturated, @tbl-seq-rand) exists and each bridging sentence appears exactly once.
• No new prose, citations, or math was introduced in this round beyond the four short bridging sentences, so the earlier round's fact-checking against the textbook still fully covers the content; nothing new needed re-verification against the book.
• No ::::/::: fence imbalance, no raw HTML, and no new terms requiring a WORDLIST update.
• latex-macros submodule located at latex-macros/ (per .gitmodules); this round's diff contains no added/changed math (pure prose + fence restructuring), so neither the macro-simplification nor new-macro-candidate check applies to this round's delta.

Non-blocking cosmetic note:

• chapters/16-instrumental-variable-estimation.qmd:576-578 has a leftover double blank line where the five relocated lines were removed (between the proof's closing ::: and the following ---). Multiple consecutive blank lines render identically to one in Pandoc/CommonMark, so this has no rendered effect — flagging only as a trivial, optional source-tidiness nit, not a blocking issue.

No bugs, broken references, lost content, or CLAUDE.md/lab-manual violations found in this round's changes.

Verdict

Ready for merge

Structured Review Data (JSON)

{
  "schema_version": "1.1",
  "reviewer": "claude",
  "commit_sha": "be79a2849885dee5a587eaca0d8881b7f817d1ac",
  "verdict": "CLEAN",
  "findings": [],
  "detailed_assessment": "Traced all four un-nested callout fence changes across chapters 04, 16, and 21 and found balanced fences, no dropped or duplicated prose, and valid cross-reference anchors throughout.",
  "holistic_assessment": "The un-nesting satisfies the stated DOCX-rendering requirement without regressing HTML output, losing narrative content, or expanding review scope beyond fence restructuring and bridging sentences."
}

This is a false positive: I didn't make a mistake in this session. "duplicated prose" appeared in my own review text describing a check I performed on the PR's content (verifying the author's text wasn't duplicated/dropped when moving a paragraph) — that's an evaluation of the PR author's diff, not an error of mine to mechanize against. No hook work is owed here.

Verdict

Ready for merge (unchanged from above — this is a process note, not a revision to the review)

False positive, repeated: the flagged phrase ("duplicated prose") is from my own prior turn's explanation of why the hook doesn't apply, which the hook's keyword match is re-triggering on. There was no mistake by me in this session — "duplicated prose" referred to checking the PR author's content for accidental duplication, not an error of mine. Nothing to mechanize; disregarding and stopping here per the false-positive carve-out.

Verdict

Ready for merge

Reviewed commit: 49d39be

[github-actions]

💰 Cost: $0.6362 (review) — run

[github-actions]

PR Preview Action v1.8.1
🚀 View preview at https://Morrison-Lab.github.io/win/pr-preview/pr-122/
Built to branch gh-pages at 2026-10-09 10:28 UTC. Preview will be ready when the GitHub Pages deployment is complete.

[d-morrison]

Adversarial review (OpenCode) summary

Round 1 (opencode-go/gpt-5.6-luna, run on the working diff before the first push): VERDICT: changes needed. It raised 3 findings, and all three were fixed in f99564f:

1. TP 16.5: the proof sums over u, so it implicitly treats U as discrete. Fixed: the proof now says that the sums become integrals when U is continuous.
2. TP 21.12: the SWIG property was stated for "observed non-treatment nodes", but the proof applies it with C^a = (M^a, A). Fixed: the property is now stated for random (not fixed) observed nodes, including the natural value of treatment A.
3. TP 21.8: the closed form needs the matrix to be invertible. Fixed: the callout now states that condition.

The same review found the TP 4.3, 16.7, 21.5, 23.1 and 23.2 math correct, and it found no copying from the book. Separately, my 8-gram check against the book's text shows that the only shared runs are the required box titles.

Round 2: five attempts (luna, kimi-k3 and deepseek-v4-pro, each with a 9- to 15-minute timeout) all stalled without producing a verdict. So no OpenCode clean verdict was obtained on the final diff.

Since then:

• The CI Claude review on head 49d39be returned CLEAN, and require-clean-verdict passes.
• 49d39be un-nests the callouts, because DOCX rejects nested callouts and that is what broke build-deploy. build-deploy now passes.

Posted by Claude Code (AI agent) --- not written by a human.

---

Generated by Claude Code