The Publications section

The ORCID/CrossRef pipeline, the manifest's fields and their traps, the cards, the metric badges and the Scholar button.

Publications

publications/publications.js over publications_manifest.json, styled in css/14-publications.css. Two tabs (List, Gallery), a multi-term search with a keyword dropdown, a sort bar, the shared pager, and a cite modal on <body>.

The list is the whole ORCID profile plus EXTRA_DOIS, minus OMIT_DOIS — 63 entries, 9 of them preprints, 54 with a figure, 52 with a PDF.

One work, one entry

A preprint and its published version must never both be in the list — the journal version replaces the preprint, it does not join it. The rule had been enforced by comparing re.sub(r"\s+", " ", title.lower()), which is not a key, it is a title with its spaces tidied, and by the time anyone counted there were four duplicate pairs standing in a list of 67. _title_key, _is_erratum and _abstract_key in update_publications.py are what replaced it. Each existed because of a specific failure:

A zero-width character is not whitespace. \s does not match U+FEFF, and Behavior Research Methods deposited "Check your outliers\ufeff!" — so the preprint sat beside its own journal version for a year while the rule that was supposed to catch it compared two strings that differed by an invisible character. The key now keeps nothing but letters and digits.
An erratum carries the paper's title plus a pointer to it. "In Medio Stat Virtus: … (vol 85, pg 1613, 2021)" is a correction notice, and CrossRef types it journal-article like anything else. ERRATUM_TAIL strips the tail, which is what makes it collide with the paper it corrects; it is dropped and the paper is kept.
A retitled preprint shares neither title nor slug with what it became. This is what most of OMIT_DOIS is: 10.31234/osf.io/jz6yq ("…Should I Use? A Data-driven Answer") against 10.1111/psyp.70164 ("…Should I Use for Psychophysiological Research?"), and 10.31234/osf.io/9pjx5 ("The Illusion Game: A Novel Experimental Paradigm…") against 10.1038/S41598-023-33148-5 ("A novel visual illusion paradigm…"), which share no opening at all — one leads with the task's name, the other with the finding. The tell there is the abstract, which survives a retitling where a title does not.

What is removed automatically and what is only reported is a deliberate asymmetry. The three rules above cannot be wrong: an identical key, an erratum tail, an abstract 90% identical to a published one. Everything short of that is reported — POSSIBLE DUPLICATE at the end of a run, naming both DOIs and what to do — because deleting a real paper because two titles happen to overlap is a far worse failure than leaving a duplicate on a page, and the hard part was never the fix. It was noticing.

The measured gap does not support automating it either way. Of the three genuine preprint pairs, the title overlaps were 1.00, 0.80 and 0.48; the closest non-duplicate pair — the Mint scale against the Affective Style Questionnaire, two questionnaire validations — was 0.27. The report fires at 0.40 and a human decides.

Two pairs still need OMIT_DOIS, and both are instructive. 10.31234/osf.io/p342w (The Heart can Lie) was retitled and reworded: its abstract is 0.77 similar to the published one, under the 0.90 a removal needs. 10.31234/osf.io/rw39q (Beauty is in the Eye of the Beholder → The Beauty and the Self) could not have been caught by anything: the titles share less than half their words and CrossRef holds no abstract at all for the published version. Its abstract was copied onto 2024_TheBeautyAndTheSelf's info.json by hand before the preprint entry went, or the site would have lost the only copy it had. Check for that before omitting a preprint.

MAX_PUBLICATIONS in update_publications.py was 20 while the section was being built and is now None. Four things follow from that number, none of which bit at twenty:

A full run is minutes, not seconds, because every DOI is cross-validated against CrossRef one at a time. It prints a heartbeat every ten works for that reason: without one it is indistinguishable from a hang, and killing it half way leaves the manifest unwritten.
Four words of a title stopped being a unique folder name. unique_slug lengthens the slug with the title's own next word until it is unique, which keeps the name content-derived and therefore stable between runs. It matters because pub_dir.mkdir(exist_ok=True) means a collision is not an error — the second publication simply overwrites the first and goes missing from a list nobody counts by hand.
ORCID is incomplete as well as occasionally wrong, and nothing reports an absence. A paper never added to the profile is simply invisible here — the Ribot centenary paper (10.3917/bupsy.549.0163) was on the old Hugo site and on no ORCID record, so it had gone missing without a trace. EXTRA_DOIS is the mirror of OMIT_DOIS: a DOI is the whole entry, fetched from CrossRef in full. An entry that later appears on the profile is skipped with a note rather than duplicated.
A folder name is a URL path, so a slug is folded to ASCII. \w is Unicode-aware in Python 3, and the first French title produced 2017_CentenaireRibotPremièrePartie — a directory that must be percent-encoded in every <img src> built from it, and that git and Dropbox normalise differently on Windows and macOS (NFC vs NFD), which is how a repository ends up with two folders whose names look identical. Every other folder is ASCII by accident of the titles being English; title_to_slug makes it so on purpose.
ORCID records are self-claimed, and "D. Makowski" is not a unique name. A hardware paper — "Novel Digital Camera with the PCIe Interface" — is on the profile with no DOI, so there was nothing to cross-validate it against and it rendered as a bare title in the middle of a psychology list. OMIT_TITLES is OMIT_DOIS for the works that cannot be named by DOI. Check any new entry that arrives with no authors: that is what an unmatched work looks like.
A work can reach the list with no year. ORCID had none for the misophonia paper, which put it in a 0000_ folder and printed "n.d." on the card, for a paper CrossRef dates perfectly well — so the year falls back to CrossRef's published/issued. ORCID's own year is never overridden. The fallback runs before the sort and before the slug, so both pick it up.

Entries carry fields written by hand in their info.json, and that is the supported way both to fix one and to add something upstream does not have: load_publications merges the file over the fetched data, so a hand-written field survives every future run — including a key ORCID and CrossRef know nothing about, which is copied through to the manifest with everything else.

github and spotify are that second kind. Both are optional, both are a plain URL, and both render as a badge on the card in publications.js beside the Altmetric donut, the Dimensions circle and the PDF — nothing generates or validates them, so a typo is a dead badge and there is no run that will say so. Neither is written into the generated page, on the same reasoning: they are about the work rather than part of it, and the page's job is the text a crawler indexes. Four badges is the widest a card gets and the row does not wrap, so a fifth link field needs the layout looked at before it is added.

2017_HowVirtualEmbodimentAffects — no DOI at all; authors taken from the old Hugo site's front matter.
2023_DisentanglingTheSocialFrom — a book chapter in Future Cities — City Futures (TU Delft OPEN Books). 10.34641/mg.55 is the book's DOI, not the chapter's, and is in neither CrossRef nor DataCite, so nothing could be fetched. Authors and venue come from the Sussex figshare deposit (sussex.figshare.com/articles/chapter/…/23737158), whose public API answers where the web page returns 403. It has no citation count for the same reason.
2025_TooBeautifulToBeFake, 2024_TheHeartCanLie and 2023_NovelVisualIllusionParadigm — spotify, a podcast episode discussing the paper. Store the bare episode URL: Spotify's share button appends a ?si= token that identifies the account that shared it, and that does not belong in a committed public link. Each of these three also has a matching .mp3 in its folder, carried over from the old Hugo site, which nothing on the new site references yet.

The publication panel

.pub-reader in publications.js, styled at the foot of css/14-publications.css. One publication, opened over the section by pressing a card's title, its figure or a gallery tile, and closed on the ✕, the backdrop or Escape.

It exists because the 63 pages generate_pages.py writes were addresses nothing on the site pointed at. A card's title linked straight to doi.org, so the only thing this site had to say about a paper was somebody else's address for it — and the pages were not merely unlinked, they were reachable: they are in sitemap.xml, so a reader arriving on /publications/<folder>/ from a search result got the shell, no handler claimed pub-<folder>, and they were left at the top of the homepage looking at the door. The route existed in routes.js and in the generator and in nothing between them.

It is the News reader and the People profile panel — same geometry, same 0.38s, same backdrop, copied rather than shared for the reason css/13-news.css already gives. Seven things are specific to it:

The DOI has not gone anywhere; it is the first action inside the panel. --primary is filled because the publisher's record is the canonical one and usually the reason somebody opened the panel at all. PDF, Code and Podcast follow it — which is where github and spotify finally get a second home, though they still stay off the generated page for the reason given above: they are about the work rather than part of it.
"In brief": the summary and the figure, in a tinted box. Two problems and one element. The figure had the panel's full width — on a 920px panel that is a 700px chart standing over two sentences of text, and at that size the figure reads as the point of the page, which it is not. And the summary is the most worth reading of the panel's prose while looking like the least of it: three lines of body text above 200 words of the publisher's abstract. So the two sit side by side in the section's green, under a heading that names them. Five things:
- The figure is a 15rem column on the right, the arrangement and the reasoning of the card the reader pressed to get here: a publication's figure illustrates an entry the title has already announced, so it follows the text rather than preceding it. Source order is visual order.
- --split is set in JS, not by :has(). 9 of the 63 entries have no figure, and their summary must take the whole row rather than the 1fr a missing figure would leave. The module knows which case it is in.
- max-height with object-fit: contain, not a fixed aspect ratio. These are charts of every shape — 667×1000 and 1000×857 both occur — and one tall enough to run past the summary beside it would put the panel back where it started. The letterboxing is invisible against the frame's white, and the frame is opaque white rather than the 72% the panel's other cards use: on the tint a translucent frame goes green too.
- The tint was measured, not picked. At the 10% it started on, the box was rgb(228, 250, 241) against the citation box's rgb(243, 250, 244) — five levels of red between two boxes that say different things. It is 16% now, rgb(228, 247, 233), fifteen levels off the citation box and seventeen off the panel's cream, and the summary's ink still clears 12:1 on it.
- The figure is a button that opens shared/media-lightbox.js, the viewer the Memories tab and the profile panel already use. A 220px chart is something a reader can only tell apart from another chart, so shrinking it without giving the full size back would have been a straight loss. That puts a third layer over the panel, which is why the Escape handler bails on defaultPrevented — the viewer claims the key from the capture phase, as people.js already documents.
The panel's four section labels are full --pub-ink, and that is a correction. "In brief", "Abstract", "Cite this" and "See also" were mixed to 75% of it, which measured 3.6:1 on all three grounds they sit on — cream, the citation box, the brief box. At 0.68rem/700 these are small text by WCAG's reckoning (large starts at 18.66px bold; these render at 10.9px), so 3.6 is a fail rather than a near miss. Undiluted they are 6.2–6.5:1. A letter-spaced uppercase label is the last thing on a page that can afford to be faint.
The abstract is fetched per publication, because the manifest deliberately does not carry it (see below). Cached in a Map, so a reader who opens the same paper twice asks once, and a failure is silent and returns {} — 36 of the 63 have no abstract anyway, so "no abstract" is an ordinary state rather than an error worth showing.
The panel opens first and the abstract lands into a slot. Waiting on the fetch before showing anything reads as a dead press. The guard on the way back in (reader.dataset.pub !== pub.folder) is what makes hopping through three "see also" links safe: a response for a paper the reader has already left must not write itself into the panel.
Re-rendering would re-roll the suggestions, so the abstract fills a slot rather than triggering a second renderPublication. That is the whole reason the render returns the empty <div>.
The card title, the card figure and the gallery tile are real anchors, via hrefForRoute — middle-click, "copy link address" and a link a crawler can follow all give the same URL writeRoute would put in the address bar. The press is still handled in script, or the browser would reload the page to reach a panel.
One Escape handler for three layers, innermost first. The image viewer (9000) is over the cite modal (1000) is over the panel (200), and a single press that closed more than one would take the reader two steps back for one keystroke. The viewer marks the event itself; the modal and the panel are ordered by hand in the one handler.

"See also" is three publications, and the keywords are what make them worth reading. A candidate is scored by how many keywords it shares with the paper being read, and the three are drawn from the best-scoring band that can fill them. Two things follow:

The panel's three are random within a band and the generated page's three are not. Opening the same paper twice should offer something new — but a page whose internal links changed on every build would churn 63 files a deploy and never let a link settle. related_publications in generate_pages.py therefore has total tie-breaks: shared keywords, then the nearest year, then the folder name. The News reader's "Keep reading" has the same shape but no scoring at all — see news.md for why.
An entry whose title key matches is dropped, not only the paper itself. This is a backstop and not the plan: a preprint and its journal version must never both be in the list in the first place, and "One work, one entry" above is where that is enforced. It stays because it is two lines, because the pipeline runs against a live ORCID profile that can hand it a pair it has never seen, and because "see also: this same paper" is the one suggestion that reads as a bug to every reader who sees it. It does not hide anything the run would have reported — the POSSIBLE DUPLICATE warning fires on the same data, in the place a maintainer will see it.

The three outgoing links are the point on the generated page. Before them a publication page's only internal links were its breadcrumbs, so 63 of the site's ~250 pages were leaves: a crawler that walks links rather than sitemaps reached one and had nowhere to go but back up. The page also carries the APA reference in the raw HTML — apa_reference in generate_pages.py, deliberately a second copy of _apaCite in publications.js, because it is the most copied string on one of these pages and both a crawler and the reader looking at the panel want it. If those two drift, that is the pair to look at.

Abstract, summary, keywords

abstract is fetched; summary and keywords are written by hand. keywords can be fetched — _crossref_keywords reads whatever CrossRef deposited — and in practice that was 5 distinct terms across 2 of the 63 entries, which is not a vocabulary. See the keyword note below.

abstract comes free from a response the script already had. Every DOI is cross-validated against CrossRef for the type check, the authors and the citation count, and the abstract was in that same message being discarded. Coverage is 27 of 63, and it is not random: the PsyArXiv preprints mostly have one, all 12 JOSS papers do not, Elsevier and Springer mostly do not, Frontiers / MDPI / Wiley / SAGE / PLOS do. An entry without one is normal. One of the 27 is hand-written — see the OMIT_DOIS note above.
It is stored as plain text and must be inserted as text, never as HTML. CrossRef returns JATS XML, and _clean_abstract strips every tag rather than translating it — unescaping entities between two stripping passes, so doubly encoded markup cannot survive. This is the only string on the site fetched from a third party; everything else (content in a post.json, summary/details in the people manifest) is lab-authored and reviewed, which is the assumption normalizeRichHtml is written on. The tag pattern requires a letter after the <, so an abstract containing "p < .05 and n > 30" keeps it.
The one exception, and it is not HTML: abstract and summary are read in the Information section's inline dialect — [label](href), **strong**, *emphasis* — by appendRichText in the panel and by inline_markup in generate_pages.py. That is shared/rich-text.js, which builds DOM nodes and only ever emits <a>, <em> and <strong>, so the paragraph above still holds: a < arriving from CrossRef cannot open an element. Four things follow.
- It exists because a deposit's own formatting arrives as punctuation. cogmod's abstract was deposited with the package name in Markdown code ticks, which rendered as literal backticks in a paragraph of serif prose — the one thing on the page that looked like a typo rather than a style. It is **cogmod** in its info.json now. _clean_abstract cannot fix this upstream: the ticks are in the deposited text, not in the JATS. The same habit reaches a hand-written field — 2021_Performance's summary had check_model() in ticks — so this is a thing to look for in a summary written about a piece of software, not only in a deposit.
- The two renderers must agree, and nothing checks that they do. INLINE_PATTERN is duplicated between shared/rich-text.js and generate_pages.py for the reason apa_reference is duplicated, and with the same consequence: if they drift, the panel and the page say different things about the same paragraph. Copy the regex, do not re-derive it.
- A machine-read copy is flattened, not rendered. strip_markup keeps the label and drops the punctuation for the JSON-LD abstract and for the <meta name="description"> — otherwise a search result prints the asterisks. The cost of the whole arrangement is that a literal asterisk in a deposited abstract would now read as emphasis. None of the 63 has one; the $\mu$ and {brms} that do turn up in a deposit are outside this grammar and pass through untouched.
- A card's search index is read back off the parser, not stripped by a second regex. plain() in publications.js builds the nodes and takes the textContent, so card.dataset.search holds exactly the string the reader sees — no asterisk they never saw, and no href behind a link's label. A search for check_model matches performance because of it.
summary carries the dialect in four places and abstract in two, which is the thing to know before adding markup to one. A summary is shown on the card, in the panel's "In brief", as the generated page's lede and as its <meta name="description"> — and the lede is <strong> in full, so a **…** inside it renders there and changes nothing. That is not a bug to fix: the alternative is printing the asterisks on the one page a crawler indexes.
A URL the deposit dropped is ours to put back. cogmod's abstract ends "openly available at ." in CrossRef itself — the preprint was deposited that way, so there is nothing to re-fetch and nothing that will ever repair it. The link was written into info.json by hand, pointing where research/creations-content.js already points for that package. Check the tail of a software paper's abstract for this: it is the sentence most likely to have lost its link, and merged.update(existing) means a hand-written abstract survives every future run.
summary is written for all 63, and it is the highest-value text an entry carries. Two or three plain sentences on what the paper found are the one thing about it that is not already on the publisher's site, on PubMed and on ResearchGate — an abstract makes a page longer, this is what makes it worth indexing. It is shown four times over: clamped to three lines on the card, in full above the abstract in the panel, as the generated page's bold lede and as its <meta name="description">.

Where there was no abstract to work from, the summary says what the paper does, not what it found. 36 of the 63 have no abstract — every JOSS paper, most of the early Springer and Elsevier entries — and an invented result is worse than a plain description. Keep that rule when adding one: a reader can tell the difference, and a wrong finding on a lab's own website is the kind of error that gets quoted back.

It has no upstream source, so merged.setdefault("summary", "") leaves whatever is in info.json alone on every run — there is nothing that can overwrite one.
keywords is a controlled vocabulary, assigned by hand to all 63. It had to be: CrossRef's own deposits gave 5 terms across 2 entries, and the keyword dropdown, the search chips and "see also" all read this one field. The vocabulary is 32 terms and it is the lab's strands rather than each paper's own words — Reality Beliefs, Interoception, Visual Illusions, Emotion Regulation, Software Development, Computational Models, Aesthetics, Self-Reference, Episodic Memory, Psychophysiology and so on, 2–4 per entry. Two constraints, neither enforced:
- A term earns its place by joining two papers. A keyword on one entry is a dropdown row that filters to a list of one, which is a worse way of finding that paper than typing its title. Time Perception is the one term currently on a single entry, and the check to run before adding another is simply to count.
- The written form is the displayed form. The dropdown, the chips and the generated page all print the string as stored; matching is toLowerCase()d, so Reality Beliefs and reality beliefs filter alike but show differently. Title Case throughout.
- info.json is the record, not this list: "existing wins" below is what carries them through every future ORCID run. There is no script that re-applies them, on purpose — a second source of truth would drift from the first.
"Existing wins" is qualified for all three. merged.update(existing) is what lets a hand-written value survive every run, but on its own it also freezes an empty one: a paper whose abstract CrossRef did not have on the first run would keep "" for ever, even after the publisher deposits it. So existing wins only when it holds something.

The manifest carries no abstract — the same rule the news pipeline follows for post bodies, and for the same reason. Every visitor downloads publications_manifest.json to render a list of titles, and 28 abstracts at a median of 165 words is ~50KB that nothing in the list view shows. Read it off info.json instead, where it costs the page nothing. summary and keywords are in the manifest: both are short and both are worth having in the list view and in the search.

Folders are not deleted when a slug changes, because a stale one may hold a PDF or a figure that was put there by hand and the new folder will not have it. The script lists them at the end of a run instead, with whatever they contain besides info.json.

Folder names are chosen, not only generated. The four-words-of-the-title rule truncates mid-phrase and keeps the filler word it lands on (…PriorsIn, …OutliersAn, TheBeautyAndThe), and renders a software paper as its subtitle rather than as the name of the thing (ModelbasedAnRPackage). SLUG_OVERRIDES replaces 45 of them. Three things about it:

It is keyed by the generated slug, not by DOI — the readable choice for a hand-edited map, and the failure is visible rather than silent: a changed upstream title stops the key matching, and the run reports both the unused override and the orphaned folder.
Two keys have to be tried, the plain four-word slug and the collision-lengthened one, because title_to_slug never returns the latter. The stale-override warning is what catches a key that matches neither.
Every override value is reserved before any slug is assigned, so a generated name that collides with a chosen one is lengthened instead — the deliberate name wins whichever order the two arrive in.

pdf and featured are the one pair of fields where disk beats info.json, and it is the opposite of every other field. They are not metadata; they are paths built out of the folder's own name, so the moment a folder is renamed the remembered value points at a file that no longer exists — existing or detected pins it there permanently. Renaming 48 folders broke 27 of 68 entries exactly this way, and the card renders a broken image. It is detected or existing now, the existing half being what lets a hand-written path to a PDF hosted elsewhere survive a run that finds no local file.

Forty of the figures came from the old Hugo site via tools/import_publication_assets.py — see the Assets table.

The section is green. --pub-accent is --section-publications (#55cc77, what it has on the hero's arc and on the brain), --pub-ink is that hue taken down to #23663a so it can be read as text on cream — #55cc77 itself is far too light for a tab label. Every tint in the sheet is color-mixed off those two, so recolouring the section is one token. It was rgba(85, 100, 160, …), a slate blue-violet that appeared nowhere else on the site, which made the tab bar, the chips and the sort buttons read as belonging to a different page. This is the arrangement News already had, and the reason both sections have three tokens rather than one.

Two things that must not take the accent:

The preprint state. A green state chip inside a green section says nothing, so --pub-preprint is the site's amber. The tint stays on the card's border, its badge, and a tint of the field faint enough that a row of mixed cards still reads as one set.

The badge sits beside the year, in .pub-card__eyebrow, and used to be position: absolute in the card's top-right corner. That corner is where the figure column is on 54 of the 63 cards, so on all but nine of them the label sat on top of a chart. Three things came out of moving it:

It is next to the year because the year is the card's other piece of bibliographic state, and that line already exists.
align-items: baseline, not center: the year is letter-spaced caps with no box and the badge is a pill with its own padding, so centring the two boxes puts the two sets of letters on visibly different lines.
The padding-right: 4.75rem the title needed to dodge the corner is gone, and with it the :not(.pub-card--has-image) qualifier that scoped it. A long title on a preprint gets its full measure back. A preprint card's eyebrow is 15.7px against a plain card's 11px — the pill genuinely is taller, and that is the whole of the difference.
The PDF and GitHub badges. Acrobat red and GitHub purple are what those things look like everywhere a reader has ever met them. They are not section colours and must not become them.

The Altmetric and Dimensions badges are placed on every card and armed only on the page the reader is looking at. Every card is built up front and the pager shows five, so writing altmetric-embed and __dimensions_badge_embed__ into the card builder handed both vendors' scripts every publication at once: 111 image requests to badges.altmetric.com and badge.dimensions.ai on the homepage, before anyone had scrolled near the section, for 61 cards that are hidden. Measured. It is 7 now. Three things:

The class name is the switch, not a data- attribute — because the class is what both scripts select on, so an unarmed badge is invisible to them by construction rather than by our remembering to filter something. The placeholders are built as .pub-metric-badge with data-metric-badge, and armMetricBadges adds the real names.
It has to be idempotent. renderView runs on every filter, every sort and every page turn, and a card already armed has already been populated — so re-arming would ask both vendors for a number they have drawn. It returns whether anything changed, and the refresh is skipped when nothing did.
ensureOfficialMetricBadgeScripts refreshed twice — once in its own .then and once in the caller's — and both vendors answer a second scan with a second round of API calls, measured at three per DOI where one was due. The inner one is gone; the refresh belongs to the caller.

The step deliberately not taken is deferring the two scripts behind an IntersectionObserver on the section. It would take the homepage to zero, at the cost of the failure mode this file already warns about for the Scholar rotation: an observer that never fires is a real case, and there it costs a timer nobody sees, while here it would leave the badges permanently blank.

The cards are plain white (rgba(255, 255, 255, 0.72)), the same field the News rows and the Creations tiles sit on. They carried an SVG fibre grain for a while — a paper texture, inline as a data: URI. Two things came out of that and are worth keeping even though the texture is gone:

The first attempt was invisible, and the opacity was the wrong knob. feTurbulence writes noise into all four channels, so its own alpha is noise centred on 0.5: pass that through and the rect's opacity is halved before anything reaches the screen, the colour is a flat mid-grey, and the result lands within a couple of levels of the background whatever the opacity says. Measured, sd 1.4 out of 255. Raising the opacity mostly darkens the card without adding contrast; what adds contrast is throwing the noise's own colour away — pin RGB with an feColorMatrix and build alpha out of the noise's luminance instead.
A texture can be measured without being seen. Draw the data URI onto a canvas over the intended background-color and take the standard deviation of the luminance. That is how both the invisible version (1.4) and the visible one (6.4) were checked in a pane that cannot screenshot.

The visible version was then rejected on sight: a page whose sections are all clean cream and white has nothing for a grain to belong to, and it read as dirt. Not a third attempt.

The figure is a column, not a thumbnail — 13rem on the right, filling whatever height the row turns out to be, so a short entry and a long one line their figures up. It used to be a 160px image centred in that space. Two things follow:

It is on the right, where News puts its thumbnail on the left, and that is not an inconsistency: a post's picture is the hook, while a publication's figure is a chart that illustrates an entry the title has already announced. So it follows the text rather than preceding it, and publications.js appends it after the body — source order matches visual order, and nothing has to reorder anything.
13rem against News' 15rem, for the same reason it is second: it is one of several things in the row rather than the row itself.

Below 780px the card stacks and the figure becomes a full-width 11rem band under the text — the same breakpoint and treatment as .news-card. The preprint badge's title padding is scoped away from cards with a figure (:not(.pub-card--has-image)): there the badge sits over the figure's column, not over the end of the title's first line.

The Scholar button is the bibliometrics, and then a joke about them. buildScholarMetrics in publications.js cycles h-index → citations → publications → "What does it mean?" → "Nothing, but check out our Google Scholar profile →", one state every SCHOLAR_ROTATE_MS (1200) — brisk on purpose, because the sequence is five states and at a leisurely pace a reader glancing at the header never reaches the punchline. Five things about it:

The numbers are hand-kept, in the SCHOLAR constant at the top of the module: there is no API for them and a Scholar profile cannot be fetched from a browser. publications: null means "use the manifest's own length", which is the honest fallback until someone writes the real Scholar figure down — the manifest is a curated selection, not the whole profile.
It is built over the anchor already in index.html, not in place of it, so the href, the target and the rel survive untouched and the no-JS fallback is a working Scholar link.
The states are stacked in one grid cell, not swapped in and out. The button is then sized once by the widest and the tallest of them and never changes size; a header that resized itself every 3.6s would shove the tab bar down the page under someone reading it. Same trick as the zoom's ask line. It is also what makes the narrow-screen rule safe: below 620px the button gives up its fixed 15.5rem and takes the row, and the punchline wrapping to three lines still does not make it move.
The rotation starts, and the observer stops it — not the other way round. It only runs while the section is on screen (an IntersectionObserver rooted on #main-page, because that is the scroll container), but an observer that never fires is a real case, and the two failure modes are not equal: rotating off-screen costs a timer nobody sees, while waiting for a callback that never comes leaves the button frozen on its first frame, which reads as broken.
Under reduced motion the three figures stand side by side and the two punchlines are dropped (.pub-scholar--still). The rotation is the effect — there is nothing to slow down — and a punchline only works after the set-up it no longer has.

FilesExpand file tree

publications.md

Latest commit

History