Session 4 — Capstone and thesis deep pages, Research, Leadership

• Date: 2026-09-13 · Branch: s4-research-leadership · PR: #7

(redacted for publication, 2026-10-01 and 2026-10-02)

Goal

IMPLEMENTATION_PLAN §4 Session 4 and its 2026-09-13 amendment, on top of Session 3a: write the capstone and thesis case studies with recreated or original figures (nothing cropped from a source PDF), ship the thesis PDF with original page 2 removed, build /research and /leadership from content, and reuse Session 3a's EmbedFrame for the valedictorian video. S0-D was not needed and the StimMap3D page was not touched.

What shipped

• Capstone body (content/projects/spinal-curvature-capstone.mdx): the problem, the origin story from SFU's article, the system as built (garment, noise-filtering firmware, live heatmap on a CAD torso), standards, Best Overall Project at ICAMES 2022, and the ML model and ultrasound brace as proposals only. Two original diagrams (the brace-fitting loop; the sensing pipeline). The team is credited and SFU's article is linked; no teammate is named on the site (see Decisions).
• Thesis body (content/projects/interpretable-medical-imaging-thesis.mdx): the question, induced ground truth (watermarks, not annotations), method (ResNet-34, VGG-16, five CAM methods, IoU at ten thresholds, paired t-tests), results with two failure cases, why it matters, and the supervisor credit. Three recreated figures: an original drawing of the ground-truth procedure and two line charts replotting the thesis's Tables 4.6 and 4.7 digit for digit. Six new frontmatter metrics (thresholds, images averaged, alpha, the two peak IoUs) each with the thesis page in its note.
• Figure pipeline. Plotted values and diagram steps live in content/figures.ts, validated by IouTableSchema / DiagramSchema and reached through figures, iouTable() and diagram() in lib/content/load.ts; their sources join allSourceIds(). components/figures/DiagramFigure.tsx (HTML flow, numbered markers, across the column from md up for a short linear flow) and components/figures/IouChartFigure.tsx (SVG lines with non-scaling strokes, every label in HTML so text keeps its size at 390 px, one accent for FullGrad and a neutral ramp with dash patterns for the rest, a legend, and the full table under a disclosure) are registered in lib/mdx.tsx as Diagram and IouChart, taking an id and a figure number.
• Thesis PDF. public/docs/paige-rattenberry-honours-thesis-2022.pdf, derived by scripts/derive-thesis-pdf.mjs (pdf-lib 1.17.1) from the staged original with page 2 removed: 49 pages to 48, page 2 of the copy is the abstract. Recorded in content/assets.json with the original's SHA-256, rights, permission, removedPages: [2], people and the text check. After the PR review (see "Review follow-up"), the script also drops the named destination to page 2, deletes every object unreachable from the trailer and renumbers the page labels, with a fixed modification date so the output is reproducible.
• Links to documents. LinkSchema.url accepts a /docs/<file>.pdf site path beside absolute URLs; lib/content/links.ts makes the loaders refuse one that is not a document in assets.json. The thesis project and research item link the derived PDF.
• Research page (app/research/page.tsx): every content/research.ts item newest first with its kind, organisation, summary through MetricStat, skills and links; items with a deep page link to it. Conference reviewing (the MLADS item, alsoResearch: true in leadership.ts) is listed under "Reviewing" from leadership data, not retyped. The thesis item now declares the project's "five" metric.
• Leadership page (app/leadership/page.tsx): every content/leadership.ts item newest first; the valedictorian item's YouTube watch link becomes a click-to-load EmbedFrame whose iframe source is derived by lib/embeds.ts (youtube-nocookie.com/embed/<id>?start=2727) while the always-visible link keeps the watch URL with t=2727s. No poster, no photos.
• EmbedFrame gained optional href (link target, default src) and allow (default fullscreen), and an embed-neutral fallback sentence. Existing tests unchanged and green.
• Deep page: FrontmatterOnly and its "written in Session 4" line are gone; the loader refuses a deepPage project without a body. DatedEntry (components/ui/) is the shared dated-list layout for Experience, Research and Leadership.
• Tests (13 files, 91; 92 after the review follow-up): deep pages have bodies and only StimMap3D has the aside; research projectSlugs resolve and repeat the project's title, dates and periods; the thesis item carries "five"; the peak metrics agree with the replotted tables; /docs/ links resolve to listed documents; no asset comes from the capstone decks or the SFU article; any thesis asset is the page-2-removed derivative; shipped PDFs have their recorded page count and, when the staged original is present, exactly the original's pages minus the removed ones; both bodies render every metric once with a public source label and numbered figures with no <img>; the capstone page says "Neither was built"; the two charts render legends, tables and peak notes; /research and /leadership render from content with the embed mounting only on click; youtubeEmbed parsing; EmbedFrame href/allow/fallback; body figure numbers run 1..n after any cover.
• LinkSchema gained an optional note (a sentence shown beside a link or its embed); the valedictorian link carries the "full ceremony, opened at the address" caption, so the page holds no fact of its own. newestFirst() in lib/content/load.ts orders both dated lists.
• Screenshots: docs/screenshots/s4-*.png (26).

Decisions and why

• Teammates are not named on the site, which Paige settled on 2026-09-15 (below). The page credits "a six-person team" and links SFU's article.
• The orthotist is "an orthotist" / "our industry collaborator".
• The thesis supervisor is named in the body, as content/research.ts already did from Session 2. Committee members are not named in page text; they appear in the shipped PDF's own acknowledgements, which assets.json records under people.
• Recreated, not cropped. The thesis's figure images are JPEG 2000 streams that nothing on this machine decodes (no pdftoppm, sharp without OpenJPEG), so the heatmap figures were not even viewed; DESIGN §3.5 forbids inventing saliency maps anyway. The two charts replot the two tables the thesis prints twice (4.1/4.6 and 4.5/4.7), the diagram draws the procedure §3.2–3.7 describe in prose. The capstone diagrams follow the architecture in the decks' own diagrams (viewed, since their images are JPEG) but are drawn from scratch with the site's tokens.
• Chart design. DESIGN §4.2 allows a neutral ramp plus the one accent, so identity comes from dash pattern and a legend rather than five hues; FullGrad, the finding, takes the accent and the only markers. Static, like a figure in a paper: no hover layer (the values sit in the table under the plot). Y scale runs to 0.7 so the peak note in the top-right corner never touches a line at 390 px; the first draft put the note beside the point and it collided.
• Figure numbers are string attributes (number="2") because this MDX compile drops JSX expression attributes (see "What the AI got wrong").
• No aside on these pages: it always appends the "How this site was built" link. Team credit and provenance are in the bodies; a test pins that only StimMap3D has an aside.
• Site-path links stay in url rather than a second field, so every existing reader keeps working; the loader, not the schema, checks that the document is listed, the way cover is resolved.
• The capstone research item links the case study; only the thesis item links a PDF. The SFU article link stays on the project (facts column) rather than being retyped.
• pdf-lib is a devDependency now (Session 6 planned it): the derivation script and the document test use it, and it appears in no page bundle.

Tools and model

Claude Code, model Claude Fable 5.1 (claude-fable-5-1). Skills: frontend-design and dataviz before the figures and page layouts; /code-review at high effort on the working tree before the PR (see "What the AI got wrong"); Playwright for axe, the projects spec, the screenshots and a scratch script that captured every figure in both themes at 1280 and 390 px; pdftotext for every citation (page by page) and for the derived PDF's text check; pdf-lib for the derivation and for a scratch image dump used to view the capstone decks' architecture slides.

Review follow-up (2026-09-14): Claude Code, model Claude Opus 5 (claude-opus-5), ran /code-review (high) on PR #7 and implemented the fixes below; pdf-lib scratch scripts inspected the shipped PDF's object graph, and a Playwright scratch script measured and captured the loop diagram.

Second review (2026-09-15): Claude Code, model Claude Opus 5 (claude-opus-5), ran /code-review (high) on PR #7 at 38ba7bd and implemented the fixes under "Second review" below.

Constraints, human decisions and implementation direction

• Task: IMPLEMENTATION_PLAN §4 Session 4 with its 2026-09-13 amendment; SESSION_PROMPTS Session 4.
• Constraints honoured: only INVENTORY-publishable files used as sources; nothing cropped from a PDF; the decks and the SFU article are not republished; the proposal PDF (mojibake, unrenderable here) was not used; page 2 of the thesis never ships; no third-party name in page text but the supervisor's; no repo links; StimMap3D untouched; the private source folder unstaged.
• Human decisions applied: the 2026-09-12 page-2 decision; the split and run order; "five" and "six-person" as metrics (the amendment); the 45:27 timestamp and the "graduand address within the full ceremony" framing (INVENTORY).
• Consent for the people thanked by name in the thesis PDF's acknowledgements: recorded in the inventory, 2026-09-15.
• Implementation direction: server components, plain data to the one client island, numbers only through MetricStat, facts only from content/, figure values in content/.
• Later sessions amended (requested by Paige, 2026-09-15, after the second review). IMPLEMENTATION_PLAN §4 Sessions 3b, 5, 6, 7 and 8, their prompts in SESSION_PROMPTS.md and AGENTS.md's architecture notes now record what this session leaves them, each checked against the code. 3b: a cover makes body figures start at 2; MDX drops expression attributes; EmbedFrame's href, allow and fixed margin. 5: researchId and projectSlug pairs are one piece of work each; order by periods; reviewing is leadership data; DatedEntry heading levels. 6: pdf-lib is present; the resume needs a document record that passes documents.test.ts; no test extracts PDF text yet, and the thesis text's reviewed phone-pattern matches need narrow exceptions. 7: chart class weight; the new pages in the sweep; one possibly flaky tooltip test. 8: public/docs/ joins the asset reconciliation.

Generated work and rejected suggestions

• Generated by Claude Code: everything in this PR, including both bodies, the figure data and components, the pages, the derivation script, the asset record and the tests.
• Rejected during the session: cropping or extracting thesis figure images (forbidden, and the JPEG 2000 streams could not be decoded anyway); an SVG-only chart (text shrank to ~7 px at 390 px, so labels moved to HTML); markers on every point of every series (noise); the peak label beside the point (collided with lines); naming the teammates (DESIGN §3.4); a separate path field for document links (every reader would change); a second dated-entry layout for Research and Leadership (DatedEntry is shared with Experience instead).
• Left undecided by this session, settled afterwards: the GenAI literature review PDF. The card said the PDF was "a Session 4 decision", but Session 4's task list covered only the thesis, and no decision was made or recorded. On 2026-09-15 Paige decided it ships in Session 6, whole, linked from its content/research.ts item, with no deep page. (Session 6b later gave it one.)
• Settled afterwards: Paige decided on 2026-09-15 that the teammates stay unnamed.

Verification (command, result, PR/commit)

• npm run typecheck, npm run lint, npm run format:check: clean.
• npm test: 13 files, 91 tests, green (claims gate over both bodies' <MetricStat> references, a body for every deep page, research projectSlug resolution, document links, assets and privacy gates over the new files and this log).
• npm run build: clean; 15 routes, all static; the three /projects/<slug> pages SSG.
• grep -l for the private source folder's path prefix over .next/server/app/**/*.html, *.rsc and .next/static/chunks/*.js: no matches.
• npm run e2e:a11y: 97 passed, both themes (twice; a third run after the review fixes had one failure, the StimMap3D tooltip test on a page this session did not touch, which passed when re-run alone, so it is load flakiness under parallel workers). tests/e2e/projects.spec.ts: 7 passed.
• $env:SESSION = "4"; npm run e2e:screenshots: 26 files.
• node scripts/derive-thesis-pdf.mjs: 49 pages to 48; pdftotext on the copy: page 1 title page, page 2 abstract, no "[ADD " text; the DESIGN §3.4 phone pattern matches only IoU table rows and DOI strings in the extracted text.
• git status shows nothing from the private source folder.
• Commits: 6bf8ba5 (content layer), 3e44645 (thesis PDF), 4cee725 (case studies and figures), 84cf214 (Research, Leadership, EmbedFrame), 653519a (tests), 8f283dc (review fixes), then this log and the screenshots (356c123). PR #7. IDs after the branch rebuild below.
• Review follow-up (2026-09-14): npm run lint, npm run typecheck: clean. npm test: 13 files, 92 tests (the new orphaned-object test and the page-label assertion both failed against the previous PDF and pass against the re-derived one). npm run build: clean, 15 routes. npm run e2e:a11y: 97 passed. tests/e2e/projects.spec.ts: 7 passed. Running node scripts/derive-thesis-pdf.mjs twice gives identical bytes; the copy has 48 page objects for 48 pages, 666 objects all reachable, no PTEX.FileName entry, and labels i, iii…viii, 1…. pdftotext (Git Bash's /mingw64/bin) on the re-derived copy: page 1 title page, page 2 abstract, extracted text identical to the previous copy's; no "[ADD " text. (The follow-up first recorded pdftotext as unavailable, having looked only on the PowerShell path.) The capstone screenshots were re-captured (SESSION=4, -g capstone); the loop note's left edge measures 131.85 px against the step titles' 131.84 px at 1280 px.
• Follow-up commits: fc66b4d (PDF derivation, test, asset record, cites), e895559 (loop-note indent and capstone screenshots), ecc6df9 (this log), 17967fb (text check), then the branch-rebuild note.
• Branch rebuild (2026-09-15, Paige's choice among ranked options, run by Paige from a script Claude Code wrote, because Claude Code's auto-mode safety check blocks force-pushes): the branch was rebuilt so the commit that first ships the thesis PDF already carries the re-derived copy. The script checked that the final tree matched the previous tip except for one dropped build-log note before force-pushing with a lease. Commit IDs from the thesis PDF commit onward changed.
• Paige reviewed the Vercel preview on 2026-09-15 and reported it correct.
• Second review (2026-09-15): npm run lint, npm run typecheck, npm run format:check: clean. npm test: 13 files, 92 tests. npm run build: clean, 15 routes. npm run e2e:a11y: 97 passed. /research screenshots re-captured (SESSION=4, -g "/research @"); /leadership renders as before, so its screenshots stand.

What the AI got wrong and how it was caught

• MDX expression attributes were silently dropped. <Diagram number={2} /> rendered without its number while number="2" worked; the ProjectPage tests failed on "Fig. 1" and a five-case probe isolated it. The bodies use string attributes and the components coerce; the reason is noted in lib/mdx.tsx.
• PowerShell wrote a UTF-8 BOM into both MDX files while switching those attributes; caught by a byte check before the next test run and stripped.
• Prose styling leaked into the diagram list ("1. back to the start" rendered with a decimal marker) because Prose's [&_ol] utilities outrank component-layer rules; caught in the 1280 px screenshot; the list resets are utilities now.
• The chart's peak label collided with the lines and clipped at the plot's left edge at 390 px; caught in the figure screenshots; moved to the top-right corner with y headroom.
• expect.anything() used with toBe in the aside test failed on first run; rewritten as a boolean comparison.
• Two thesis claims were first drafted too strongly and corrected against the text before the body was committed: "the maps were noise" in the low-accuracy VGG-16 run (FullGrad still found the watermark, p. 29) and "significant at every threshold except the highest" as if it applied to both models (it is VGG-16's 90–95% only, p. 30).
• /code-review (high) was cut off by the session rate limit while orchestrating, as in Sessions 2 and 3a: three of its five finder angles died with the orchestrator, two (line by line, efficiency) completed and returned twelve candidates, each verified by hand against the code and the screenshots. Confirmed and fixed (commit 8f283dc):
  1. The four-step pipeline diagram rendered as four cramped columns at 390 px. The column count was an inline grid-template-columns applied at every width while only the gap was behind md:; the screenshot showed the words broken mid-syllable. The count is a custom property read only inside the md media query now, and the 390 px screenshot is vertical.
  2. A fact in a page: the "SFU's recording of the full ceremony, opened at the address" caption was JSX in app/leadership/page.tsx and would have captioned any future YouTube link the same way. It is the link's note in content/leadership.ts.
  3. Number(number) could ship "Fig. NaN" for a missing or non-numeric MDX attribute (MDX is not type-checked) and a table value past the chart's 0.7 scale would plot outside the box while the peak note still quoted it. Both components throw at build.
  4. A cover plus a body would number two figures "1", since the page renders a cover as Fig. 1 and bodies hard-code their sequence. A content test now requires body figure numbers to run consecutively from 2 when the project has a cover, 1 otherwise.
  5. key={step.title} would collide on a repeated step title (index keys now); unrounded polyline floats inflated the SVG strings that ship twice (rounded to a tenth of a unit).
  6. Tests: the shipped PDF was parsed three times (memoised); a /docs/ link test could never fail on its own because the loaders throw first (reduced to a sentinel); a loop re-rendering every deep page duplicated the per-page tests (folded in); the newest-first sort was written in both pages (newestFirst() in load.ts). Not changed, with the reason: the chart's per-node utility classes cost about 1.5–2 KB raw per chart per serialisation; the reviewer marked it optional and Session 7 owns the budget pass.

Review follow-up (2026-09-14)

/code-review (high) on PR #7 returned three findings; all were confirmed by hand and fixed.

• The removed page still shipped inside the PDF. pdf-lib's removePage() only unlinks a page from the page tree, and save() writes every object in the context. The first copy held 49 page objects for 48 pages: the old page 2 and the objects only it used, still reachable through a named destination. No viewer showed it, and the session's checks (pdftotext, the page count and page sizes) only looked at pages in the tree, so none of them could catch it. The script now drops destinations that target a removed page, deletes unreachable objects (10) and refuses to write if anything still points at a removed page. tests/unit/documents.test.ts fails on any orphaned object or on a page-object count that differs from pageCount. The assets.json transformation that said nothing else changed was corrected, with a new transformation recording the re-derivation.
• Viewer page labels were one off past page 1. /PageLabels still started arabic numbering at index 8, so the abstract showed as "ii" and printed p. 31 as "30". The labels are now rebuilt from the original's minus the removed page (i, iii…viii, 1…), tested against the original when the staged original is present. The IoU figure cites said "PDF p. 39", the original's page; the site links the copy, where it is p. 38, so the cites and the schema comment now name the shipped copy's page.
• The loop diagram's "back to the start" note was not indented. The list's [&>li]:pl-0! reset outranked the component rule. The first fix, a pl-[2.55rem]! utility on the note, still measured 0 px, because .x > li outranks a single class when both are !important; the reset is now scoped to .flow-step items.

Second review (2026-09-15)

/code-review (high) on 38ba7bd found no correctness bugs and three low findings; all were confirmed by hand and fixed.

• The log named a commit the branch rebuild had replaced. It is gone from this log; the current commit IDs are unaffected.
• my-6 on the leadership video frame did nothing. EmbedFrame always adds my-8, which comes later in the generated CSS. The class is removed rather than made to win, so the page keeps the spacing its screenshots show, and className's doc comment now says margins lose.
• The Reviewing entry's h3 outsized its section h2 on /research (text-2xl under text-xl). DatedEntry sizes an h3 text-xl, as Card and Experience's h3s already are.

Screenshots

docs/screenshots/s4-*.png: every route at 1280 px and 390 px (home in both themes); the two new deep pages, /research and /leadership are the ones this session changed. In 24 of them the footer's build hash was blanked; nothing else in any image changed.