Friday Harbor Laboratories

Definition.—A publication is counted when Friday Harbor Laboratories, or the station under one of its earlier names (Table V), appears in the publication or its record: in an author address, in the main text (methods, acknowledgments, figure captions or anywhere else except the reference list), in the station's own publication series, or in a dataset's metadata. A mention only of the town of Friday Harbor, of another Friday Harbor organization, or only in a reference list does not count. Each publication is assigned the name the station used in its publication year.

Search.—Candidates came from OpenAlex (author-affiliation strings searched for every name variant, the FHL institution record, full text, titles and abstracts); the Web of Science Core Collection (an export of — records matching addresses, funding acknowledgments and topics, run through the University of Washington Libraries under the compiler's own sign-in); PubMed and PubMed Central (affiliations and full text); Europe PMC (searches of the affiliation, methods and acknowledgment sections); Crossref (affiliation, bibliographic and funder queries); DataCite, Zenodo, Dryad, GBIF, NCBI BioProject and BCO-DMO for datasets; UW ResearchWorks and other repositories for theses and student papers; OSTI and USGS for reports; archive.org full-text scans, including the Biodiversity Heritage Library, BioStor and JSTOR Early Journal Content, for the station-era literature; FHL's own publication lists, recovered from the Internet Archive; and news coverage found by web search. Name variants were matched with patterns that tolerate hyphenation, line breaks and the British spelling "Harbour", and with the citation forms of the station's own series. The searches ran on —.

Merging.—The searches returned — rows, merged into — candidate records (— with a DOI). Two rows were joined when they shared a DOI or a PubMed, PubMed Central or OpenAlex identifier, or when their titles matched closely (a fuzzy similarity of at least 93 out of 100), their years differed by no more than one, and their first authors shared a surname.

Screening.—Unambiguous candidates were decided by written rules, such as an author address that names the laboratory, or a match that falls only in a reference list. The rest were judged by AI agents against a written screening guide that defines every field, lists the station's names and describes the common false positives: the town, the ferry, other Friday Harbor organizations (the Center for Whale Research, The Whale Museum and others), Seattle-campus oceanography, and bound journal volumes that merely contain a station paper. For each record an agent recorded whether it is FHL-connected, the kind of connection, whether FHL appears in the affiliation and in the main text (and in which sections), the publication type, its confidence, a reason and a short quotation of the evidence. — records were judged in a first round and — in a second, after the search was extended and an address classifier corrected. Calibration agents then re-judged samples across batches and changed — inconsistent calls. Of the final decisions, — rest on rules, — on AI screening, — on calibration review, — on the audit, — on de-duplication and — on the clean-up review described below.

Resolution.—The — candidates with no visible evidence were opened one at a time, reading the full text or landing page wherever it was freely available: — were confirmed, — rejected, and — could not be decided; they stay among the unverified candidates.

Where FHL appears.—"Affiliation" is an author address naming the station. "Main text" is anywhere else in the publication except the reference list. Where no full text could be read, the category says "text not checked" rather than assuming the station is absent.

Evidence tiers.—Verified: FHL was seen in the record (an author address, a passage of text, the station's own series, or dataset metadata). Probable: a full-text index names the lab, but the passage could not be read. Listed by FHL: on an FHL-maintained publication list but not independently confirmed; the audit found an FHL connection in only about a third, so these are not added to the verified count.

Audit.—A stratified random sample of — decisions, drawn from every tier and from the excluded and unverified records, was re-checked by separate AI agents told to try to prove each decision wrong from the publication itself (— read in full text). A second agent adjudicated every disagreement. Precision is the share confirmed, weighted by stratum size, with Wilson 95% intervals (Table III), and the best estimate counts each tier at its precision. The audit was done by AI agents, not by people; a human check of a sample would strengthen it.

Recall.—FHL's own publication lists serve as an independent sample of FHL work. For each decade, recall is the share of listed papers that the database searches found on their own (Fig. 19). Listed papers that the searches missed make up the "Listed by FHL" tier.

Types and citations.—Publication types come from each source's metadata, corrected by the screening agents, and are grouped into seven categories for Fig. 12. Citations are OpenAlex cited-by counts at retrieval, with Crossref, Europe PMC or DataCite counts where OpenAlex had none; other databases give different numbers. The citation rate in Fig. 16 divides a paper's citations by its age in years, counting the year of publication as one.

Research topics.—For Figs. 2 to 6, AI agents gave every included publication one of 13 research topics, an optional second topic and one of 13 organism groups, working from its title, journal and, for — of the —, an abstract. They followed a written guide with definitions, rules for borderline cases and worked examples. The publications were shuffled into random batches, so that each batch's mix of topics could be checked against the whole. A separate agent labelled a random sample of — publications without seeing the other labels: the two agreed on the topic for —% of them (Cohen's κ —) and on the organism group for —% (κ —). Because both labellings come from the same model following the same guide, this measures consistency, not correctness. Of the — random batches, — matched the overall mix of topics within the range expected by chance. A final review applied three consistency rules to — labels: disease, symbiosis and microbiome studies take the host's organism group; research on people goes to Other (or to Whiteley Center, below); and fossil and taphonomy work, for which the guide had no topic, was gathered into a fourteenth topic, Paleobiology & taphonomy (docs/topic_guide.md). Publications whose only link to FHL is the Whiteley Center, the station's study center for visiting scholars and writers, are labelled Whiteley Center, with their field as sub-topic (Fig. 4): — publications, — of them probable, inferred from a full-text match on non-marine work published since 2000.

Clean-up.—A final review, made after the topic classification, merged — duplicate records into —. These were the same article under two DOIs, a preprint and its published version, repository and reprint copies, and the same dataset deposited twice. The kept record takes the strongest FHL evidence and the highest citation count in its group; a dataset, a conference abstract or a thesis counts as a work of its own. The review also removed — records that are not publications or whose only FHL match lies in a reference list (whole issues, newsletters, front matter, other institutions' catalogues and reports, meeting-abstract volumes, peer-review documents, data downloads), and moved — obituaries, biographies and histories to Mention only, as the screening guide prescribes. The audit's error rates were measured before this review, which removed many of the errors it found, so the best estimate is now somewhat conservative.

People.—FHL scientists are the authors who list FHL, or the station under an earlier name, as their address on at least three verified publications, identified through OpenAlex's author records. For each, OpenAlex's history of the institutions on their papers gives where they came from (the earliest institution with a paper before or in their first FHL year) and where they went (the institution with papers in the most years after their last FHL paper, counting only institutions with at least two such years); those still publishing from FHL in the last three years count as still at FHL. Citation links count how often the FHL papers of one scientist cite the FHL papers of another, from OpenAlex reference lists. Advisor-student links come from UW thesis records, which name the advisor, and from Wikidata's doctoral-advisor and doctoral-student statements. People were matched to Wikidata by ORCID or by Wikidata's own links to the station, and advisors were followed up to five generations back; a match was dropped if it would make the person younger than 18 at their first FHL paper. Mentors were also inferred from co-authorship. Among the verified publications (datasets, news items and papers with more than 25 authors left out), a scientist who is last author on at least two of a colleague's first-author papers is taken as that colleague's mentor, if they lead a group (last author for at least two different first authors), began publishing at least four years earlier, and the pair's first paper comes within eight years of the colleague's first publication. Careers come from OpenAlex affiliation histories, and people are matched by OpenAlex author ID or by surname and given name. The rule found — links. It recovers — of the — documented advisor-student pairs that appear in the census with the student first and the advisor last; most of the rest share only one paper. AI agents then checked every inferred link against public records (theses, lab pages, CVs), with an independent second look at every verdict that did not rest on a direct source and an adjudicator where the two looks disagreed: — of the — were advisor, postdoctoral or student-research mentors and are drawn as documented, — were collaborations between peers and were removed, and — could not be settled and are drawn as inferred (Table VI). Lab rosters add — documented links: AI agents read each FHL scientist's public lab, team and alumni pages and CVs and listed the graduate students, postdocs, technicians, interns and staff named there, and a second agent re-opened every page to confirm each entry. Undergraduates named on a home-university lab page are included only if they also appear as authors in the census, and the Star Lab team comes from FHL's own pages, read from Internet Archive copies. Graduate committees listed in CVs add — links, drawn dotted, and advisors and mentors named in the fact-checked profiles (below) add —. Every FHL scientist was searched, in three rounds (the last starting from each profile's sources and ORCID record, since web-search capacity ran short in the first two): — have a roster, CV list or committee list; — lead no lab of their own (students, early-career researchers or staff); — worked before public lab pages and no list of their students is public; and for — who taught, no public list of trainees was found. Each FHL scientist and each teacher in the network has a two-paragraph profile (— in all), shown when the person is picked in "Find a person": an AI agent wrote it from the person's census record and public sources (faculty and lab pages, CVs, ORCID, obituaries), and a second agent checked every claim against those sources and revised — of them. In all the network holds — links among — people. OpenAlex groups papers into people automatically and sometimes merges or splits people with similar names. Two such merges were corrected by hand: L. D. Partridge (FHL papers 1975–79) and Volker Schmid (1988–92) had each been merged with a different scientist of the same name, a biogerontologist and a statistician. Scientists from before about 1970 are under-represented because older papers carry no address metadata.

The chronicle.—Sources for section X were gathered through public interfaces: newspaper pages from Chronicling America (Library of Congress) that mention the station, mostly the San Juan Islander of 1904–1914; FHL's annual newsletter Intertidal Tidings and its history and news pages, read from Internet Archive copies; the Historical Centennial Timeline compiled by Claudia Mills and Colin Hermans (used for facts, credited, not quoted at length); the census's news items; and catalogue records of the University of Washington Libraries' photographs. AI agents extracted — dated events from these, another merged them into one timeline, and four more checked each event against its sources, correcting — (Table VI); the timeline now holds — events, — of them added later from the records of newly found images. A second run looked for images: eight agents each searched one kind of source (Wikimedia Commons and Wikidata; openly licensed photographs through Openverse; publications from before 1931 on the Internet Archive; University of Washington Libraries items that are free to reuse, among them the Board of Regents' minutes, yearbooks and trade journals; Chronicling America newspaper pages, cut to the story; USGS and NOAA maps and NAIP aerial imagery; federal and Library of Congress collections; and sources for 1930–1999). They found — candidates; a second agent opened each image and its record, confirmed the licence and wrote the caption, passing —; a picture editor chose — for balance across the eras; and fact-checkers checked every caption, credit and licence again. Images (— in all) are shown only where they are in the public domain (published in the United States before 1931, works of the US government, or marked so by the holding library) or openly licensed (CC0, CC BY, CC BY-SA). The UW Libraries' photographs that carry an "In Copyright" statement are linked rather than reproduced, and other promising images whose rights could not be confirmed were left out.

Not searched.—Google Scholar and Scopus (no automated access), BIOSIS Previews, Zoological Record and ASFA. Before about 2008 databases did not index acknowledgments, so counts are a lower bound, especially for 1985–2010 and before 1970.

Software and files.—Python 3 scripts (pandas, rapidfuzz, pypdf, lxml, openpyxl) collect, merge, screen and summarise the records, and the charts are drawn with d3, version 7. The scripts and their run order, the screening guide, the agent instructions and every intermediate decision are kept with the census files, so each count can be traced to the records and rules behind it.

Platform.—The census, its code, the spreadsheet, the figures and this page were produced with Claude Code 2.1.295, Anthropic's coding agent, in the Claude desktop app on macOS, on 13–16 October 2026. The coordinating session and every agent it started used the model Claude Opus 5.5 (claude-opus-5-5). The census and the first version of this page were produced at the "xhigh" reasoning-effort setting, the second highest of five (low, medium, high, xhigh, max); from the topic classification onward (14 October, 18:10 UTC), including the people section and the check of inferred mentors, the setting was "max", the highest; the chronicle's second image search and a review of the page's phone layout (15–16 October) ran at "xhigh" (Table VI). Claude Code's ultracode mode was on while the census was built, so each substantive step ran as a scripted multi-agent workflow; the topic classification and the mentor check, added afterwards, ran as agents started directly. In all, — runs used — agents (Table VI). The first complete version took — hours of wall-clock time, most of it unattended.

Division of labour.—Claude planned the pipeline, wrote all the code, queried the public sources through their programming interfaces, judged ambiguous records, ran the audit, and drafted and designed this page. The compiler set the task and the later requests listed below, signed in to the UW Libraries and Web of Science with their own account, and saved the exported files; Claude operated the Web of Science export in a browser after that sign-in. No person screened individual records.

Safeguards.—Agents had no accounts or passwords, did not get around paywalls, bot checks or CAPTCHAs, and treated fetched pages as data rather than as instructions. Paywalled publications were judged from their metadata and abstract and are labelled "text not checked". Personal email addresses found in records were removed from the outputs.

Limits.—AI judgments can be wrong; the audit's error rates (Table III) and the topic agreement rates above describe how often. Treat individual records as leads to check, and the totals as estimates with the stated uncertainty.

Prompts.—The work followed from the prompts below, typed into the Claude desktop app. Messages about signing in, saving files and progress are summarised rather than quoted. Each prompt shows when it was typed (Pacific time) and the output tokens (text, reasoning and tool calls) it led to: those Claude generated in the main session before the next prompt, and those of the agents started in that time; a workflow's agents count toward the prompt during which the workflow was launched. Many agent records keep only the opening token count of each response, so agent totals are estimated from the length of what the agents wrote (about 2.3 characters per token, as measured on the main session) wherever that is larger. The agents worked from instructions that Claude wrote: a screening guide, a search specification, a topic guide and the workflow scripts, which hold every agent prompt. They are kept with the census files (docs/screening_guide.md, docs/harvest_spec.md, docs/topic_guide.md and docs/ai_workflows/).