How this site was built

A separate review of the branch or pull request before Paige's review: Claude Code's code review, a review agent, or a second Claude Code session, and once a Claude Code review of Codex's work. Automated review, not Paige's.

• The mobile contour was a crop of the desktop field again. Small review follow-up, 2026-09-13
• The Session 1 and 2 prompts and tasks were rewritten, altering the record. Small review follow-up, 2026-09-13
• Two build-log templates had diverged. Small review follow-up, 2026-09-13
• The Resume link was indented at 390 px. Small review follow-up, 2026-09-13
• A staged asset's digest was recorded but never checked. Small review follow-up, 2026-09-13
• The headshot loader's error did not say the asset must be an image. Small review follow-up, 2026-09-13
• Card copy dropped the capstone team size and the thesis method count. Small review follow-up, 2026-09-13
• StimMap3D's public commit count was recorded wrong. Session 0-D
• Checks stopped at the numbers: a block quote was attributed to a README section that does not exist. Session 0-D
• The first publication gate had gaps: a term wrapped across a line, staged names outside a folder, home-folder shapes, silent skips. Session S0-G
• The fixed gate still had gaps: long or bracketed old names, pattern terms across a line break, runs in a fork. Session S0-G
• The scrub went past removal in four places, dropping credits and decisions. Session S0-G
• Smaller scrub gaps: a dangling bullet, unmarked redactions, stale links. Session S0-G
• The link-accent utility emitted no CSS. Session 1
• The prose max-width token was never applied. Session 1
• The mobile menu's open state was keyed to the pathname. Session 1
• The contour's path data went over the wire twice. Session 1
• The end-to-end specs could run against a development server. Session 1
• outline-none removed the focus ring. Session 1
• git log ran once per route. Session 1
• Project names were typed into page copy. Session 1
• The third ink colour failed AA contrast in both themes, and axe never saw it. Session 1
• A fixed root font size overrode the reader's setting. Session 1
• Nothing enforced Prettier. Session 1
• A route count was off by one. Session 1
• The stock favicon shipped. Session 1
• A shell call through cmd.exe ate the git format string. Session 1
• The footer date was a day ahead. Session 1
• format:check would fail on a Windows clone with CRLF line endings. Session 1
• The accessibility gate skipped the 404 page and the open mobile menu. Session 1
• A stale production server could be reused by the specs. Session 1
• Home's "Download resume" button led to a placeholder page. Session 1
• The page scrolled sideways at phone width. Session 2
• A hidden tooltip widened the page. Session 2
• Bullets were not verbatim: dashes had changed. Session 2
• Metric matching had no token boundary. Session 2
• The GMLE certification printed twice. Session 2
• A deprecated image prop. Session 2
• A source label fell back to the raw source id. Session 2
• Pages imported raw content modules, skipping validation. Session 2
• StimMap3D appeared on About without its disclaimer. Session 2
• A figure caption overprinted at 768 px. Session 2
• A tooltip trigger was 11 px wide. Session 2
• The contour weighed more than its budget. Session 2
• Smaller cleanups from the review. Session 2
• Tooltips ran off phone screens. Session 2
• Tooltips failed WCAG 1.4.13 (dismissable, hoverable). Session 2
• Hero contour lines stopped short of the right edge. Session 2
• Splitting text by metrics could drop a metric. Session 2
• A private staging path was served in the /projects page. Session 3a
• Two metrics rendered without their source. Session 3a
• The case studies were reordered. Session 3a
• Three different definitions of a deep page. Session 3a
• Nested complementary landmarks. Session 3a
• A proof line carried numbers without metrics. Session 3a
• A fact typed as a page literal. Session 3a
• The EEG-TMS card named StimMap3D without its disclaimer. Session 3a
• loading="lazy" on a click-mounted iframe could outlast the load timeout. Session 3a
• Filter chips could be clicked before hydration. Session 3a
• A malformed MetricStat tag passed silently. Session 3a
• The claims gate over-counted a metric in a tagline. Session 3a
• The embed frame lost keyboard focus after loading. Session 3a
• The live region was inserted already filled. Session 3a
• A whole asset record, staging path included, was serialized into a client island's props. Session 3b
• The embed's two branches tested different conditions, so the disclaimer could fail open. Session 3b
• The aside's links bypassed both link gates. Session 3b
• A relaxed digest pattern turned a malformed revision into a skipped check. Session 3b
• A dead fallback in the capture script. Session 3b
• A cramped diagram at 390 px. Session 4
• A caption fact typed in a component. Session 4
• A figure could be numbered NaN or plot values out of scale. Session 4
• Two figures were numbered 1. Session 4
• Key collisions and long floats in the chart markup. Session 4
• A test that could never fail. Session 4
• A removed page was still inside the derived PDF. Session 4
• Page labels were off by one. Session 4
• A diagram note was not indented. Session 4
• The log named a commit that no longer existed. Session 4
• A margin utility did nothing. Session 4
• An h3 rendered larger than its h2. Session 4
• The StimMap3D timeline entry lacked its disclaimer. Session 5
• The dimming rule recoloured hover and the tooltip. Session 5
• The live region could announce a stale filter. Session 5
• The Home teaser had no size cap. Session 5
• CI never ran the explorer and projects specs. Session 5
• Teaser labels were sized wrong. Session 5
• prebuild rewrote the layout, so its guard could never fire. Session 5
• Transitions ran under reduced motion. Session 5
• Hovering under a filter dimmed the filter's own neighbourhood. Session 5
• Environment files loaded after the site URL was read. Session 6
• A stale overflow PDF survived a good build. Session 6
• No predev hook rendered the resume. Session 6
• Some strings bypassed the resume's font-run splitter. Session 6
• A document exception was itself phone-shaped. Session 6
• A NUL byte in the build log. Session 6
• The site URL was not validated. Session 6
• The resume's include flags were untested. Session 6
• Typos in regular expressions. Session 6
• Section headings were typed in the generator while AGENTS.md said no fact is typed. Session 6
• The JSON-LD escape did nothing. Session 6
• A trailing slash on the site URL failed the build. Session 6
• The resume generator ignored the production environment file. Session 6
• Definitions were attributed to the literature review's sources. Session 6b
• An entry count was wrong. Plan amendment, 2026-09-20
• A metric was sent to the wrong file. Plan amendment, 2026-09-20
• A label that fits was listed as one that does not. Plan amendment, 2026-09-20
• A change to Home's skills strip was called a reorder. Plan amendment, 2026-09-20
• An acceptance item was too broad. Plan amendment, 2026-09-20
• "Measured" covered one label, not two. Plan amendment, 2026-09-20
• The font spec skipped the 404 page. Session 7
• The local Lighthouse script could orphan its server. Session 7
• The served font family took the loader's const name, a reserved font name. Session 7
• The JS budget subtracted framework files the page did not load. Session 7
• A content-visibility rule contradicted what shipped. Session 7
• The plan called Session 7 shipped while it was in review. Session 7

Something the working agent did before committing: looked at a screenshot, measured, re-read a source or its own diff, or ran a scratch script.

• A find-and-replace swapped two headings and orphaned a prompt's body. Plan amendment, 2026-09-15
• Two cross-references said "above" for a note below. Plan amendment, 2026-09-15
• A draft run order dropped Sessions 6 and 7. Plan amendment, 2026-09-13
• A PowerShell pipe's encoding blocked Unicode replacements. Small review follow-up, 2026-09-13
• A spelled-out metric put its source button inside a hyphenated word. Small review follow-up, 2026-09-13
• A resume revision introduced claims the sources do not support. Session 0
• The same revision gave a leadership role the wrong start date. Session 0
• A resume draft ran to three pages. Session 0
• Two Session 4 tasks could not be done, and a third would have published an excluded page. Session 0
• The pull request's description was stale. Session 0
• The hero deep link pinned nothing: its preset was the default. Session 0-D
• Setup let an import create a production deployment that cannot be deleted. Session 0-E
• A smoke test pushed an already-deployed commit, and a stale status read as a pass. Session 0-E
• Changing the framework preset silently reset the Ignored Build Step. Session 0-E
• Backslashes broke two replacement scripts. Session S0-G
• A CI step lost its printf escape. Session S0-G
• Subagents dated their redaction notes with the session's start date. Session S0-G
• A rewritten log sentence read backwards. Session S0-G
• The commit-message check would have failed the first public commit. Session S0-G
• A directory change persisted across shell calls, and a write failed. Session 1
• Prettier rewrapped the planning documents. Session 1
• Prettier broke a block that had to stay verbatim. Session 1
• The Contact page showed list bullets. Session 1
• The contour ran behind the heading on mobile. Session 1
• A perl substitution mangled a test file. Session 2
• The hero failed the 30-second test. Session 2
• The small-screen field was a crop again. Session 2
• A tooltip used the wrong font. Session 2
• A certificate issuer had no source. Session 2
• "Independent project" headed the capstone page. Session 3a
• A spelled-out count sat in prose, outside the claims gate. Session 3a
• Two tests used a real repository's URL as their refusal case. Session 3b
• A 16:10 poster cropped the app's banner. Session 3b
• An element screenshot had the banner stitched in. Session 3b
• Lazy images were blank in screenshots. Session 3b
• PowerShell wrote a byte-order mark. Session 4
• Prose styles leaked into a diagram's list. Session 4
• A chart's peak label collided. Session 4
• Two thesis claims were drafted too strongly. Session 4
• Constellation labels piled up. Session 5
• A drafted amendment said there was no link field; there was. Session 5
• A literal NUL byte served as a key separator. Session 5
• The fix for it wrote the wrong text, twice. Session 5
• A header comment named the wrong path. Session 5
• The resume's page number never rendered. Session 6
• A tool decoded escapes into a NUL byte. Session 6
• The favicon came out black. Session 6
• Heredocs were rejected. Session 6
• The resume's provenance revision was stale. Session 6
• Two of three old label references were missed. Session 6b
• Three sentences the literature review does not support. Session 6b
• A module comment was false. Session 6b
• A phrase was looser than the resume's wording. Session 6b
• The review pass was narrower than asked. Session 6b
• A heredoc failed, and a script assumed Python. Session 6b
• A label change was said to need no layout regeneration. Plan amendment, 2026-09-20
• The headline was said to appear nowhere else. Plan amendment, 2026-09-20
• Grid classes that do not exist were named, and two tests missed. Plan amendment, 2026-09-20
• The new deep page was said to register nothing. Plan amendment, 2026-09-20
• A heredoc hit a file-name length limit. Plan amendment, 2026-09-20
• An inline script broke on an apostrophe. Plan amendment, 2026-09-20
• Thesis chart labels ran together at 360 px. Session 7
• A video's title ran under its button at 360 px. Session 7
• A Lighthouse aggregation called median-run was not the median. Session 7
• CI kept no Lighthouse reports. Session 7
• The extras fonts sat after Inter in the stack. Session 7
• content-visibility on the timeline nearly shipped, and broke anchor links. Session 7
• The plan still named the old font loader. Session 7

A Vitest test other than the claims and privacy gates.

• The contour grid overshot the view box. Session 1
• A fallback waited for an iframe error event that never fires. Session 3a
• A test fixture retyped research fields. Session 3a
• Grid cards rendered only the tagline, so a ranking went unrendered while the claims gate passed. Session 3a
• Heredocs halved the backslashes in a regular expression. Session 3b
• MDX expression attributes were silently dropped. Session 4
• A matcher was used where it cannot match. Session 4
• A slug equal to a research id broke a test. Session 5
• The literature review PDF had no title. Session 6
• The resume failed the orphan-object check, twice. Session 6
• A pairing was typed twice. Session 6b
• The plan's task list was incomplete in three places. Session 6b
• The font test expected a name record the subsetter drops. Session 7

A Playwright test against the production build, other than axe.

• Two concurrent Playwright runs shared one results folder. Small review follow-up, 2026-09-13
• A test tried to focus a heading. Session 2
• A locator counted nested list items. Session 3a
• A live region sat in the server-rendered fallback. Session 5
• A pattern lost its backslash, and the screenshots timed out. Session 5
• A "flaky" tooltip test was smooth scrolling. Session 5
• A metadata test over every page timed out. Session 6
• A PDF check counted page objects in raw bytes. Session 6
• Pages below the root lost their Open Graph image. Session 6
• The tooltip test's wait raced smooth scrolling. Session 7
• A Lighthouse flakiness estimate was wrong within the hour. Session 7

tsc, next build, the resume generator's page guard, or a failed deployment.

• The first Vercel build ran with no framework preset. Session 0-E
• Typecheck failed on a clean tree. Session 1
• Heredoc writes never landed. Session 2
• A Required<> type broke. Session 2
• __dirname pointed elsewhere under the bundler. Session 3a
• A click handler sat in a server-rendered fallback. Session 3a
• Git Bash's path rewriting broke the build. Session 5
• The resume overflowed to three pages, then four. Session 6
• The PDF renderer could not be required by the script runner. Session 6

The phone pattern over every text file and shipped PDF, the staged-asset checks, and since S0-G the publication gate.

• Hex checksum digits matched the phone pattern. Small review follow-up, 2026-09-13
• Heredocs halved the backslashes in the new gate's own tests. Session S0-G
• A staged revision broke the digest check. Session 3b
• A random-number divisor matched the phone pattern. Session 5
• A document exception was too narrow. Session 6
• A log draft quoted a document identifier fragment. Session 6
• A log draft held a CI run id, which is phone-shaped. Session 7

The tests that every number is a sourced metric, present in its own text.

• About's no-digits check tripped on real names. Session 2
• Thesis metrics were declared but never rendered. Session 2

ESLint, including the React Compiler rules.

• Two state updates inside effects. Session 1
• A state update inside an effect. Session 5

Axe over every page in both themes.

• Dimming entries by opacity failed contrast. Session 5

Something Paige asked or noticed. Her review at merge is not written into the logs: only Session 1's log records her verdicts, so this group counts what the logs say, not what she caught.

• The README's status had not changed since Session 2. Session 5