Search UK Biobank's openly published field catalogue — 11,821 variables in 410 categories with participant counts, units, the instanced/arrayed shape that decides how many columns a field becomes, and the 172 retired fields the website hides but the download still carries — and report what an access application requires, applications being paused as of August 2026. Never fetches participant data.

UK Biobank field catalogue

UK Biobank followed ~500,000 UK adults recruited 2006–2010, and has kept adding to them: biochemistry, NMR metabolomics, Olink proteomics, exomes, genomes, imaging, accelerometry, and linked hospital, cancer and death records.

The whole data dictionary is an open download. No account, no application. That answers the question researchers actually ask before applying — is what I need even in there, and on how many people? — in seconds rather than an afternoon of clicking a web interface.

This skill never touches participant data, and cannot get you any. It reads the catalogue: field names, units, categories, participant counts, dates, and the structural metadata that says how a field is shaped. Individual-level values are controlled and require an approved application, a signed material transfer agreement, and a fee — see Requesting access, which describes that route and does not promise it.

Read this before you plan a study

Two facts settle most triage questions, and both cut the opposite way from the usual controlled-access resource:

Consent is broad, not disease-specific. The Access Procedures state the resource is available to all bona fide researchers "for all types of health-related research that is in the public interest", with no preferential or exclusive access. There is no per-study consent restriction to check the way there is on a disease-specific cohort. What binds instead is health-related and public interest.
You do not need your own ethics approval for a data application. UK Biobank is registered as a Research Tissue Bank and its clearance covers data and sample applications; the Access Procedures say separate REC or other ethical clearance "is not required nor needed". Re-contact applications are the exception and do need their own.

So the binding constraint here is usually not consent. It is whether the measurement exists at a useful sample size — which is exactly what the catalogue answers for free.

Downloading the catalogue

One CGI endpoint serves every schema table as TSV. Schema 1 is the field catalogue.

curl -sS "https://biobank.ndph.ox.ac.uk/showcase/scdown.cgi?fmt=txt&id=1" -o primary.tsv
curl -sS "https://biobank.ctsu.ox.ac.uk/crystal/scdown.cgi?fmt=txt&id=1" -o mirror.tsv
wc -c primary.tsv mirror.tsv
shasum -a 256 primary.tsv mirror.tsv
cut -f1,2,13,24 primary.tsv | head -3

Printed 2026-08-18:

 4164390 primary.tsv
 4164390 mirror.tsv
 8328780 total
2d7b5f394457002a5a41e278074d64c3b56811f4c50fd916599145aa10210cc6  primary.tsv
2d7b5f394457002a5a41e278074d64c3b56811f4c50fd916599145aa10210cc6  mirror.tsv
field_id	title	units	num_participants
3	Verbal interview duration	seconds	501109
4	Biometrics duration	seconds	497769

Same bytes, same digest, so the crystal mirror is a drop-in fallback when the primary is down. A third path, biobank.ndph.ox.ac.uk/ukb/scdown.cgi, served the same file.

scdown.cgi answers HTTP 200 with an HTML error page when it does not recognise the table id — Content-Type: text/html, body "Sorry, internal error prevents download of schema". Fed to a TSV parser that trusts the status code, it yields ~40 rows whose sole column is named <!DOCTYPE HTML>, with no exception raised. codown.cgi behaves the same way. Every loader below checks the first bytes rather than the status.

The tables worth knowing, all at scdown.cgi?fmt=txt&id=<n>:

id	table	rows on 2026-08-18	why you want it
1	data field properties	11,821	the catalogue itself
3	categories	410	`main_category` is an integer — this names it
9	instancing dictionaries	12	what an instance index means for a field
2	encoding dictionaries	858	which coding a categorical field uses
13	category browse tree	362	parent/child, for walking a category subtree

Individual codings come from a different script, codown.cgi?id=<encoding_id>, and the full index of schema tables is listed at showcase/schema.cgi.

The 29 columns, and the code tables

Schema 1's header is:

field_id title availability stability private value_type base_type item_type strata
instanced arrayed sexed units main_category encoding_id instance_id instance_min
instance_max array_min array_max notes debut version num_participants item_count
showcase_order cost_do cost_on cost_sc

Almost everything interesting is an integer code. The mappings are published, but only as prose on https://biobank.ndph.ox.ac.uk/showcase/schema.cgi?id=1 — there is no machine-readable download of them, so hardcode the table and re-read that page when a code you have never seen appears:

column	codes
`availability`	0 full · 1 view-only · 4 retired
`stability`	0 complete · 1 updateable · 2 accruing · 3 variable
`private`	0 public · 1 restricted
`value_type`	11 integer · 21 categorical single · 22 categorical multiple · 31 continuous · 41 text · 51 date · 61 time · 101 compound · 201 blob
`base_type`	0 see value type · 11 integer · 31 real · 41 string · 51 date
`item_type`	0 data · 10 sample · 20 bulk · 30 record
`strata`	0 primary · 2 auxiliary · 3 derived
`instanced`	0 no · 1 defined · 2 variable
`arrayed`	0 no · 1 yes
`sexed`	0 unisex · 1 male · 2 female

Three of those need more than a lookup:

value_type is the shape of one value. On the 2026-08-18 snapshot: 4,951 continuous, 3,698 categorical-single, 1,345 date, 1,008 integer, 533 text, 171 categorical-multiple, 85 time, 30 compound. No 201 blob field existed, though the code is documented.

base_type is only meaningful for categorical fields, which the data proves rather than the documentation. Cross-tabulating the two columns, base_type is 0 for every one of the 7,952 non-categorical fields, and non-zero for exactly the 3,869 categorical ones — 3,818 coded as integers and 51 as strings (single choice splits 3,664 integer / 34 string, multiple choice 154 / 17). So for a categorical field, base_type tells you whether the codes behind the labels are integers or strings, which is what you need to join against the coding in schema 2. For anything else it carries no information.

item_type tells you whether the field is even a column in your extract, and this is the one that quietly breaks analysis plans. 0 data (11,271 fields) lands in the tabular dataset. 20 bulk (424) is a separate per-participant file download — DICOMs, VCFs, accelerometer records. 30 record (112) lives in a Data Portal record table, not in the tabular extract at all. 10 sample (14) refers to physical material. A plan that assumes every field is a column will mis-size 550 of them.

The last three columns are fee tiers, and only one of them is populated. cost_do is 0 for all 11,821 fields. cost_on — the tier for analysis on the Research Analysis Platform — is 1 for 11,266 fields, 2 for 413, 3 for 141, and 0 for exactly one, field 33 Date of birth. The showcase page renders those as o1/o2/o3 and the lone zero as oX, so treat 0 as no tier stated, not as tier zero — mapping it to an integer tier is how a one-field hole becomes a wrong fee. The rest tracks the published three-tier fee structure: the 141 tier-3 fields are exome and whole-genome sequencing and their QC metrics, and tier 2 is dominated by imaging-derived measures, polygenic risk scores and cardiac monitoring. So the open catalogue tells you which fee tier your shortlist sits in before you ask. cost_sc renders as s<N> (and as sX when it is 0), and its definition is not published in the schema documentation — that page still documents a single tier column where the file now has three. Do not infer a meaning for cost_sc; report the raw value.

Searching and grouping — the triage layer

Match on title and category name, group by category, lead with the participant count, and carry availability. Grouping matters because a title search alone scatters related fields, and the category is often where the useful name lives. Carrying availability matters for the reason in the next section: the download contains 172 retired fields that the Showcase website does not show you.

import csv, io, re, urllib.request
from collections import defaultdict

SHOWCASE = "https://biobank.ndph.ox.ac.uk/showcase"

def load(sid, encoding):
    raw = urllib.request.urlopen(f"{SHOWCASE}/scdown.cgi?fmt=txt&id={sid}", timeout=300).read()
    # scdown.cgi answers HTTP 200 with an HTML error page for an unknown table id.
    if raw.lstrip()[:9].lower().startswith(b"<!doctype"):
        raise SystemExit(f"schema {sid}: server returned HTML, not TSV — check the table id")
    return list(csv.DictReader(io.StringIO(raw.decode(encoding)),
                              delimiter="\t", quoting=csv.QUOTE_NONE))

fields = load(1, "utf-8")     # data-field properties
cats   = load(3, "cp1252")    # category id -> title
cat_name = {c["category_id"]: c["title"] for c in cats}
MARK = {"4": " RETIRED", "1": " view-only"}   # availability; "0" is full

def triage(query, min_n=0):
    """Fields whose title or category matches `query`, grouped by category, reporting the
    participant count that decides whether the study is powered — and whether the field is
    still live. Retired fields stay in the download but vanish from the Showcase website."""
    pat = re.compile(query, re.I)
    groups = defaultdict(list)
    for f in fields:
        cname = cat_name.get(f["main_category"], f"[{f['main_category']} unknown]")
        if pat.search(f["title"]) or pat.search(cname):
            n = int(f["num_participants"])
            if n >= min_n:                      # -1 and 0 fall out at min_n=1
                groups[(f["main_category"], cname)].append(
                    (f["field_id"], f["title"], n, f["availability"]))
    print(f"query {query!r}  min_n={min_n}")
    for (cid, cname), hits in sorted(groups.items(), key=lambda kv: -max(h[2] for h in kv[1])):
        ns = [h[2] for h in hits]
        dead = sum(1 for h in hits if h[3] == "4")
        print(f"  cat {cid:>6}  {cname[:38]:38} {len(hits):>4} fields  "
              f"n {min(ns):,}-{max(ns):,}" + (f"  [{dead} retired]" if dead else ""))
        for fid, title, n, avail in sorted(hits, key=lambda h: -h[2])[:3]:
            print(f"            {fid:>8}  {title[:44]:44} n={n:,}{MARK.get(avail, '')}")

triage(r"\bliver\b", min_n=10000)
print()
triage(r"liver iron|townsend")

Printed 2026-08-18:

query '\\bliver\\b'  min_n=10000
  cat   1039  Food (and other) preferences              1 fields  n 181,941-181,941
               20686  Liking for liver                             n=181,941
  cat    126  Liver MRI                                 7 fields  n 10,094-94,790
               20204  Liver Imaging - T1 ShMoLLI - DICOM           n=94,790
               20254  Liver imaging - IDEAL protocol - DICOM       n=86,699
               40063  Acquisition protocol                         n=41,641
  cat    149  Abdominal composition                     1 fields  n 47,036-47,036
               24352  FR liver PDFF mean                           n=47,036
  cat    158  Abdominal organ composition               5 fields  n 10,069-38,967
               21080  Liver volume                                 n=38,967
               21088  Liver PDFF (fat fraction)                    n=30,736
               21089  Liver iron                                   n=30,736
  cat   2411  Digestive system disorders                2 fields  n 17,707-17,707
              131670  Date K76 first reported (other diseases of l n=17,707
              131671  Source of report of K76 (other diseases of l n=17,707
  cat   1000  [1000 unknown]                            2 fields  n 15,112-15,116
             1020254  EMBARGOED UNLINKED : Liver imaging - IDEAL p n=15,116
             1020204  EMBARGOED UNLINKED : Liver Imaging - T1 ShMo n=15,112

query 'liver iron|townsend'  min_n=0
  cat 100094  Baseline characteristics                  2 fields  n 501,315-501,315  [1 retired]
                 189  Townsend deprivation index at recruitment    n=501,315 RETIRED
               22189  Townsend deprivation index at recruitment    n=501,315
  cat    126  Liver MRI                                 4 fields  n 1,110-40,354  [2 retired]
               40060  Liver iron (Fe)                              n=40,354
               40062  Liver iron corrected T1 (ct1)                n=34,291
               22417  Liver iron corrected T1 (ct1)                n=2,812 RETIRED
  cat    158  Abdominal organ composition               2 fields  n 10,069-30,736
               21089  Liver iron                                   n=30,736
               21093  Liver iron (Fe) - gradient echo              n=10,069

The first output is the whole point of the skill, and it also shows why you cannot rank by num_participants alone: the largest hit is a food-preference questionnaire item. The real answer — quantitative liver fat on ~30,700 people, liver volume on ~39,000, raw imaging on ~95,000 — is three lines down. Read the titles.

The [1000 unknown] group is the orphan-category trap arriving unannounced: a main_category that schema 3 does not define, so an inner join would have deleted those two rows silently. All 126 fields in it are bulk items marked private = 1, so their presence in a hit list is not the same as their being available. They split exactly in half — 63 titled EMBARGOED UNLINKED : and 63 EMBARGOED LINKED : — and that word is the whole question for an imaging file, because it says whether the file can be joined to a participant at all. Do not read the prefix off one of them and generalise.

Retired fields are in the download and not on the website

This is the trap that survives a careless search, because nothing in the field's own numbers looks wrong. availability = 4 marks 172 fields as retired. They are still in schema 1, with full titles, full participant counts and a live field.cgi page that says "Field is currently retired" — but the Showcase's category pages do not list them at all. So a shortlist built from the download does not match the website, and the difference is silent.

The second triage output above is the counter-example. 189 and 22189 are both titled Townsend deprivation index at recruitment, both report n=501,315, and 189 is retired. Nothing but availability separates them. Eleven retired fields carry the exact title of a live replacement, and three of those differ by an order of magnitude in n:

retired	n	current	n	title
`22400`	1,110	`40060`	40,354	Liver iron (Fe)
`22402`	4,609	`40061`	40,746	Proton density fat fraction (PDFF)
`22417`	2,812	`40062`	34,291	Liver iron corrected T1 (ct1)
`189`	501,315	`22189`	501,315	Townsend deprivation index at recruitment

Whole categories can be dead. Category 2000 "Hospital inpatient" holds 36 fields in schema 1, with num_participants up to 413,163 — and all 36 are retired, so its Showcase page lists none of them. A query on hospital inpatient|episodes containing returns 86 fields of which 86 are retired. Categories 2002–2005 are largely the same story: the summary-diagnosis and summary-operation fields were superseded by the first-occurrence and record-level tables.

The ground truth is one HTTP request away, and it is worth making once for any category you plan to build on: label.cgi?id=<category_id> lists the live fields, so fields in schema 1 − retired == fields the page lists. It holds on all 22 categories that contain a retired field — but only if you parse the listing table (<tr class="row_odd|row_even" id="f<id>">). Scraping every field.cgi?id= on the page picks up cross-references out of the category's own prose and fails on two of the 22: category 149 reports 29 against a true 28, and 150 reports 7 against 5, because its description names two retired fields in a sentence about them being retired. Category 126 Liver MRI is the compact case — 11 fields in the download, 7 on the page, difference exactly the 4 retired ones.

Note how easily this hides. The \bliver\b search above uses min_n=10000, and all four retired Liver MRI fields have n below 10,000, so they never appear. A threshold chosen for a good reason silently suppressed the evidence.

Two more columns are easy to skip for the same kind of reason. private = 1 (319 fields, including 33 Date of birth and parents' months of birth) marks fields carrying disclosure risk, not fields you can simply request. availability = 1 (54 fields, including per-chromosome genotype probabilities) is view-only — documented on the showcase, not handed out. Filter on all three before you build a shortlist.

Two habits that make this reliable. Search the category name as well as the title, because UK Biobank names the assay at the category level and the analyte at the field level. And set min_n deliberately: the imaging sub-cohort is tens of thousands, not half a million, so a min_n tuned to the full cohort silently hides every imaging field.

`instanced` and `arrayed` — the part that trips people

These two columns are why UK Biobank fields are hard to reason about, and they are independent of each other.

instanced means the field was captured more than once per participant, indexed by an instance. The index does not universally mean "visit":

instanced = 0 (4,605 fields) — captured once, almost always at instance 0. Two exceptions carry instance_min = instance_max = 3: 41000 and 41001, the COVID-19 re-imaging fields. Read the index off instance_min, not off the instanced flag — field.cgi calls both of them "Singular" and does not print an index at all.
instanced = 1 (7,198 fields) — a defined instancing scheme. instance_id points into schema 9, which says what the instances are. instance_id = 2 is the common one: initial assessment centre plus later repeat visits.
instanced = 2 (18 fields) — variable. The index has no fixed meaning across participants. Field 40000 (Date of death) uses instance_id 9000001, "Death registry reports": instance 0 and 1 are two registry reports, not two visits. Treating a variable instance as a timepoint is a real analysis error, and it is only 18 fields, so it is easy to miss.

arrayed means one participant can have many values at the same instance — a list, not a repeat measure. arrayed = 1 on 550 fields.

Total columns a field becomes in a tabular extract is the product, named <field_id>-<instance>.<array>:

import csv, io, urllib.request

SHOWCASE = "https://biobank.ndph.ox.ac.uk/showcase"

def load(sid, encoding):
    raw = urllib.request.urlopen(f"{SHOWCASE}/scdown.cgi?fmt=txt&id={sid}", timeout=300).read()
    if raw.lstrip()[:9].lower().startswith(b"<!doctype"):   # HTTP 200 + HTML error page
        raise SystemExit(f"schema {sid}: server returned HTML, not TSV — check the table id")
    return list(csv.DictReader(io.StringIO(raw.decode(encoding)),
                              delimiter="\t", quoting=csv.QUOTE_NONE))

fields = {f["field_id"]: f for f in load(1, "utf-8")}
inst   = {i["instance_id"]: " ".join(i["descript"].split()) for i in load(9, "utf-8")}

def shape(field_id):
    f = fields[field_id]
    lo, hi = int(f["instance_min"]), int(f["instance_max"])
    alo, ahi = int(f["array_min"]), int(f["array_max"])
    n_inst = (hi - lo + 1) if f["instanced"] != "0" else 1
    n_arr  = (ahi - alo + 1) if f["arrayed"] == "1" else 1
    n, items = int(f["num_participants"]), int(f["item_count"])
    print(f"{field_id}  {f['title']}")
    print(f"  instanced={f['instanced']} arrayed={f['arrayed']}"
          f"  -> {n_inst} instance(s) x {n_arr} array slot(s) = {n_inst * n_arr} columns")
    print(f"  column names: {field_id}-{lo}.{alo} ... {field_id}-{hi}.{ahi}")
    if f["instanced"] != "0":
        print(f"  instance {f['instance_id']} means: {inst.get(f['instance_id'], '?')[:96]}")
    if n > 0:
        print(f"  participants {n:,}  item_count {items:,}"
              f"  -> {items / n:.2f} values per participant with data")
    else:
        print(f"  participants {n} (sentinel — not published)  item_count {items:,}")

for fid in ["23400", "41270", "40000", "30069"]:
    shape(fid); print()

Printed 2026-08-18:

23400  Total Cholesterol
  instanced=1 arrayed=0  -> 2 instance(s) x 1 array slot(s) = 2 columns
  column names: 23400-0.0 ... 23400-1.0
  instance 2 means: All participants attended an initial assessment centre. A proportion were invited several years
  participants 488,513  item_count 507,961  -> 1.04 values per participant with data

41270  Diagnoses - ICD10
  instanced=0 arrayed=1  -> 1 instance(s) x 259 array slot(s) = 259 columns
  column names: 41270-0.0 ... 41270-0.258
  participants 448,651  item_count 7,276,575  -> 16.22 values per participant with data

40000  Date of death
  instanced=2 arrayed=0  -> 2 instance(s) x 1 array slot(s) = 2 columns
  column names: 40000-0.0 ... 40000-1.0
  instance 9000001 means: Death registry reports
  participants 56,961  item_count 57,021  -> 1.00 values per participant with data

30069  Cell images
  instanced=2 arrayed=1  -> 16 instance(s) x 104 array slot(s) = 1664 columns
  column names: 30069-0.0 ... 30069-15.103
  instance 693 means: Used to group together items related to the same case report.
  participants 1,124  item_count 19,074  -> 16.97 values per participant with data

Read those four together and the whole model falls out. 23400 is one measurement repeated at a second visit for a minority — 488,513 people, 507,961 values, so about 19,400 repeats. 41270 is not visit-indexed at all: a single instance slot holding up to 259 ICD-10 codes per person, 16 on average. 40000 is a variable instance, where the index is a registry report. 30069 combines instancing and arraying and becomes 1,664 columns for 1,124 participants — which is what a wide extract of an arrayed, instanced field looks like, and why you compute the column budget before requesting one.

item_count / num_participants is the cheapest sanity check available: near 1.0 means one value per person, well above 1 means the field is genuinely multi-valued. Both columns carry -1 on the same 270 fields and 0 on the same 12, so the ratio never silently divides by a fake count — but neither column can be summed.

The independent check on all of this is the field's own showcase page, which declares the shape in its header table: Instances reads Singular, Defined (N) or Variable (N) and Array reads No or Yes (N), alongside Participants and Item count. Those four were compared against this arithmetic on nineteen fields spanning every value_type, both item_type 20 bulk and item_type 30 record, array_min = 1, the registry-indexed instancing schemes and the retired and embargoed ends of the catalogue, and they agreed every time. One nuance the arithmetic cannot see: declared slots are not populated slots. 30900 declares Defined (4), and its page then breaks out 3 instances, because nobody was measured at instance 1. Four is still the right column budget; three is the right number of timepoints.

Resolving a field to its showcase page

field.cgi?id=<field_id> is the human-readable page. It answers HTTP 200 for field ids that do not exist, so status code is not a validity test — validate against the catalogue you already downloaded.

import csv, io, urllib.request

SHOWCASE = "https://biobank.ndph.ox.ac.uk/showcase"

raw = urllib.request.urlopen(f"{SHOWCASE}/scdown.cgi?fmt=txt&id=1", timeout=300).read()
known = {f["field_id"]: f["title"] for f in
         csv.DictReader(io.StringIO(raw.decode("utf-8")), delimiter="\t",
                        quoting=csv.QUOTE_NONE)}

def resolve(field_id):
    """field.cgi answers HTTP 200 for ids that do not exist, so validate locally."""
    field_id = str(field_id)
    if field_id not in known:
        return None, f"field {field_id} is not in the catalogue"
    return f"{SHOWCASE}/field.cgi?id={field_id}", known[field_id]

for fid in ["23400", "30900", "41270", "99999999"]:
    url, note = resolve(fid)
    print(f"{fid:>9}  {note[:44]:44} {url or ''}")

for fid in ["23400", "99999999"]:
    body = urllib.request.urlopen(f"{SHOWCASE}/field.cgi?id={fid}", timeout=60).read().decode(
        "utf-8", "replace")
    print(f"  field.cgi?id={fid:<9} HTTP 200  "
          f"'Field is not in database' present: {'Field is not in database' in body}")

Printed 2026-08-18:

    23400  Total Cholesterol                            https://biobank.ndph.ox.ac.uk/showcase/field.cgi?id=23400
    30900  Number of proteins measured                  https://biobank.ndph.ox.ac.uk/showcase/field.cgi?id=30900
    41270  Diagnoses - ICD10                            https://biobank.ndph.ox.ac.uk/showcase/field.cgi?id=41270
 99999999  field 99999999 is not in the catalogue
  field.cgi?id=23400     HTTP 200  'Field is not in database' present: False
  field.cgi?id=99999999  HTTP 200  'Field is not in database' present: True

Nor is the body string a test. field.cgi?id=99999999 does carry "Field is not in database", but field.cgi?id=0 and field.cgi?id=abc answer 200 without it. The local membership check is the only reliable one, which is why it comes first.

Categories resolve the same way at label.cgi?id=<category_id> — and that page is also the ground truth for which fields in a category are still live, since it omits retired ones.

Field and category notes contain the showcase's own cross-reference markup: ~F30900~ a field, ~C143~ a coding (not a category), ~L100116~ a label, ~R4654~ a resource, ~P1874~ a publication, ~Ehttps://…~ an external link. C meaning coding rather than category is a genuine ambiguity — coding 143 is the Olink assay list, while category 143 is "Cannabis use".

Traps confirmed on the live file

num_participants is a sentinel column, not a count. 270 fields carry -1 and 12 carry 0 on the 2026-08-18 snapshot. -1 means the count is not published — the showcase page renders those fields' participant count as the word pending, so it is not "nobody". item_count carries -1 on exactly the same 270 rows. Summing or averaging either raw column silently subtracts.
172 fields are retired and the website does not show them. availability = 4. They stay in the download with full counts; label.cgi omits them. Eleven of them carry the exact title of the live field that replaced them. See Retired fields are in the download and not on the website above — this is the single easiest way to size a study on a dead field id.
The download endpoints answer HTTP 200 with an HTML error page. Both scdown.cgi?id=<unknown> and codown.cgi?id=<unknown> return status 200 and Content-Type: text/html. A TSV parser accepts the HTML without raising and hands back rows whose first column is named <!DOCTYPE HTML>. Check the leading bytes, not the status.
The two files disagree on text encoding. Schema 1 is valid UTF-8. Schema 3 (categories) is cp1252 — decoding it as UTF-8 raises UnicodeDecodeError on byte 0x97 at offset 152,914, an em dash inside a category description. Decode each table with the encoding that table actually uses.
126 fields reference a category that does not exist. Every field's main_category resolves against schema 3 except 1000, which is absent from the category table entirely. All 126 are item_type 20 bulk items with private = 1, and they split 63 EMBARGOED UNLINKED : … / 63 EMBARGOED LINKED : … — the two prefixes are field-id twins (1020204 / 2020204), so reading one and generalising gets the linkage backwards for half of them. An inner join on category drops all 126 without a word; use an outer join and label the orphan.
availability = 7 is undocumented. Three fields carry it (12652, 12663, 12704, all brain-MRI "reason not performed" fields) and the schema page documents only 0, 1 and 4. Pass unknown codes through as raw values rather than mapping them to a guess.
codown.cgi output has no trailing newline, while scdown.cgi output does. So wc -l undercounts a coding file by one row. Use grep -c '', or parse it.
Titles and notes contain literal double-quote characters — 8,756 across schema 1, 200 of them in titles, spread over 4,083 rows. On this snapshot they are all mid-field, so a default CSV parse and quoting=QUOTE_NONE give byte-identical results (checked with both csv and pandas 2.3.2). Pass QUOTE_NONE anyway; it costs nothing and removes the failure mode entirely.
The documented column list is stale. schema.cgi?id=1 ends its list with showcase_order | tier; the file ends with showcase_order cost_do cost_on cost_sc. Read the header row, never the documentation, to order columns.
What the catalogue does not contain: any participant value, any per-field missingness pattern, which participants have data, or the distribution of a variable. num_participants is a headcount of people with at least one non-missing value across all instances — not the analysable n for a specific visit, and not a complete-case count for a model with covariates. Treat it as an upper bound.

Two worked answers

Both of these are aging-project shaped questions, and the catalogue answers them in opposite ways.

NMR metabolomics — three categories, 512 fields, ~488,500 people.

category	title	fields	max `num_participants`
220	NMR metabolomics	251	488,514
221	NMR metabolomics QC indicators	249	412,098
222	NMR metabolomics processing	12	488,521

Category 220 is the biomarker panel itself — field ids 20280–20281 and 23400–23648, all value_type 31 continuous, instanced=1 with instances 0–1 (baseline and repeat), and units 163 mmol/l, 77 percent, 4 ratio, 3 nm, 3 g/l, 1 degree. Representative ids: 23400 Total Cholesterol (n=488,513), 23407 Total Triglycerides (488,513), 23444 Omega-3 Fatty Acids (487,913), 23470 Glucose (487,597), 23478 Creatinine (479,247), 23480 Glycoprotein Acetyls (488,514).

Note the number. The published NMR biomarker atlas is titled for 118,461 individuals; the catalogue on 2026-08-18 reports ~488,500 for those same fields, roughly four times as many. The catalogue is the current answer and a paper is a historical one — which is exactly why you read num_participants rather than inheriting an n from the literature. Category 221 is the paired QC layer — all 249 titles end , QC Flag, all are value_type 21 categorical, and 76 of them carry num_participants = -1. Budget for them: a per-biomarker QC flag is how you exclude a bad spectrum, and it is a second column per analyte.

Olink proteomics — five fields, and not one protein among them.

Category 1838 "Proteomics" has a single child, category 1839 "Protein biomarkers", containing exactly five fields: 30900 Number of proteins measured (n=53,039), 30901 Plate used for sample run (53,039), 30902 Well used for sample run (53,039), 30903 UKB-PPP Consortium selected participant (6,230), 30904 Number of blind-spike duplicates (1,318).

Searching the catalogue for a protein name finds nothing, and that is not a search failure. 30900 has item_type = 30 — record — meaning the normalised protein expression values live in a Data Portal record table (olink_data), one row per participant-instance-protein, not as thousands of tabular fields. Selecting 30900 is what grants access to that table. Any plan that budgets "one column per protein" is built on a shape that does not exist.

The assay list is open, though, which answers "is my protein on the panel?" without an application. It is coding 143:

curl -sS "https://biobank.ndph.ox.ac.uk/showcase/codown.cgi?id=143" -o olink_assays.tsv
grep -c '' olink_assays.tsv
head -3 olink_assays.tsv
grep -E "^[0-9]+\s(GDF15|LEP|IL6|APOE);" olink_assays.tsv

Printed 2026-08-18 — 2,924 lines, so 2,923 assays:

2924
coding	meaning
1	A1BG;Alpha-1B-glycoprotein
2	AAMDC;Mth938 domain-containing protein
154	APOE;Apolipoprotein E
1137	GDF15;Growth/differentiation factor 15
1418	IL6;Interleukin-6
1572	LEP;Leptin

meaning is GENE;description, and the integer coding is the protein id used inside the olink_data table. Split on ; to get a gene-symbol lookup. For an organ-specific proteomic clock this file plus 30900's n=53,039 is the complete feasibility answer: which of your marker set is measurable, on how many people.

Get the files

This writes catalogue metadata to disk as CSV — field descriptions, counts and shapes. It does not download and cannot download participant data.

Set QUERY to a case-insensitive regular expression matched against field titles and category names. Every code column is decoded to a label, columns_in_extract is precomputed, the -1 sentinel is preserved as a flag rather than coerced to a number, and a manifest records the SHA-256 of each source file so a later rebuild is comparable.

import csv, hashlib, io, json, os, re, urllib.request

SHOWCASE = "https://biobank.ndph.ox.ac.uk/showcase"
OUT      = "Data/ukb_catalogue"
QUERY    = r"nmr|olink|protein biomarker"   # case-insensitive, matched on title + category

VALUE_TYPE = {"11": "integer", "21": "categorical single", "22": "categorical multiple",
              "31": "continuous", "41": "text", "51": "date", "61": "time",
              "101": "compound", "201": "blob"}
ITEM_TYPE  = {"0": "data (tabular)", "10": "sample", "20": "bulk (files)",
              "30": "record (portal table)"}
STABILITY  = {"0": "complete", "1": "updateable", "2": "accruing", "3": "variable"}
INSTANCED  = {"0": "no", "1": "defined", "2": "variable"}
AVAILABLE  = {"0": "full", "1": "view-only", "4": "retired"}

os.makedirs(OUT, exist_ok=True)
manifest = {"source": SHOWCASE, "query": QUERY, "files": [], "downloads": []}

def fetch(url, dest, encoding):
    raw = urllib.request.urlopen(url, timeout=300).read()
    if raw.lstrip()[:9].lower().startswith(b"<!doctype"):   # HTTP 200 + HTML error page
        raise SystemExit(f"{url} returned HTML, not TSV — check the table id")
    with open(dest, "wb") as fh:
        fh.write(raw)
    manifest["downloads"].append({"url": url, "path": dest, "bytes": len(raw),
                                  "sha256": hashlib.sha256(raw).hexdigest(),
                                  "encoding": encoding})
    print(f"  {len(raw):>10,} bytes  {dest}")
    return list(csv.DictReader(io.StringIO(raw.decode(encoding)),
                              delimiter="\t", quoting=csv.QUOTE_NONE))

print("downloading:")
fields = fetch(f"{SHOWCASE}/scdown.cgi?fmt=txt&id=1", f"{OUT}/schema_1_fields.tsv", "utf-8")
cats   = fetch(f"{SHOWCASE}/scdown.cgi?fmt=txt&id=3", f"{OUT}/schema_3_categories.tsv", "cp1252")
insts  = fetch(f"{SHOWCASE}/scdown.cgi?fmt=txt&id=9", f"{OUT}/schema_9_instancing.tsv", "utf-8")

cat_name = {c["category_id"]: c["title"] for c in cats}
inst_txt = {i["instance_id"]: " ".join(i["descript"].split())[:120] for i in insts}
pat = re.compile(QUERY, re.I)

rows = []
for f in fields:
    cname = cat_name.get(f["main_category"], f"[category {f['main_category']} not in schema 3]")
    if not (pat.search(f["title"]) or pat.search(cname)):
        continue
    lo, hi = int(f["instance_min"]), int(f["instance_max"])
    n_inst = (hi - lo + 1) if f["instanced"] != "0" else 1
    alo, ahi = int(f["array_min"]), int(f["array_max"])
    n_arr = (ahi - alo + 1) if f["arrayed"] == "1" else 1
    n = int(f["num_participants"])
    rows.append({
        "field_id": f["field_id"],
        "title": f["title"],
        "category_id": f["main_category"],
        "category": cname,
        "num_participants": "" if n < 0 else n,   # -1 means "not published", not zero
        "n_is_sentinel": n < 0,
        "item_count": f["item_count"],
        "units": f["units"],
        "value_type": VALUE_TYPE.get(f["value_type"], f["value_type"]),
        "item_type": ITEM_TYPE.get(f["item_type"], f["item_type"]),
        "availability": AVAILABLE.get(f["availability"], f"code {f['availability']}"),
        "stability": STABILITY.get(f["stability"], f["stability"]),
        "instanced": INSTANCED.get(f["instanced"], f["instanced"]),
        "instances": n_inst,
        "instance_meaning": inst_txt.get(f["instance_id"], "") if f["instanced"] != "0" else "",
        "array_slots": n_arr,
        "columns_in_extract": n_inst * n_arr,
        "encoding_id": f["encoding_id"],
        "rap_cost_tier": f["cost_on"],
        "debut": f["debut"][:10],
        "version": f["version"][:10],
        "showcase_url": f"{SHOWCASE}/field.cgi?id={f['field_id']}",
    })

if not rows:
    raise SystemExit(f"no field matched /{QUERY}/ — widen it. Titles name measurements, "
                     "not constructs, so a derived quantity has no field to match")

rows.sort(key=lambda r: (r["category"], int(r["field_id"])))
csv_path = f"{OUT}/ukb_fields_filtered.csv"
with open(csv_path, "w", newline="", encoding="utf-8") as fh:
    w = csv.DictWriter(fh, fieldnames=list(rows[0].keys()))
    w.writeheader()
    w.writerows(rows)

manifest["files"].append({"path": csv_path, "rows": len(rows),
                          "bytes": os.path.getsize(csv_path)})
manifest["catalogue_version"] = max(f["version"] for f in fields)[:10]
manifest["fields_in_catalogue"] = len(fields)
manifest["contains"] = ("field-level metadata only — no participant data, "
                        "no individual-level values")
with open(f"{OUT}/manifest.json", "w") as fh:
    json.dump(manifest, fh, indent=2)

print(f"\n{len(rows)} fields matched /{QUERY}/ -> {csv_path}")
print("catalogue version:", manifest["catalogue_version"])
for cid, cname in sorted({(r['category_id'], r['category']) for r in rows},
                         key=lambda t: int(t[0])):
    sub = [r for r in rows if r["category_id"] == cid]
    ns = [r["num_participants"] for r in sub if r["num_participants"] != ""]
    cols = sum(r["columns_in_extract"] for r in sub)
    dead = sum(r["availability"] == "retired" for r in sub)
    print(f"  cat {cid:>5}  {cname[:34]:34} {len(sub):>4} fields  "
          f"max n {max(ns) if ns else 0:>7,}  {cols:>5} extract columns"
          + (f"  {dead} RETIRED" if dead else ""))

Printed 2026-08-18:

downloading:
   4,164,390 bytes  Data/ukb_catalogue/schema_1_fields.tsv
     176,100 bytes  Data/ukb_catalogue/schema_3_categories.tsv
       1,530 bytes  Data/ukb_catalogue/schema_9_instancing.tsv

517 fields matched /nmr|olink|protein biomarker/ -> Data/ukb_catalogue/ukb_fields_filtered.csv
catalogue version: 2025-09-20
  cat   220  NMR metabolomics                    251 fields  max n 488,514    502 extract columns
  cat   221  NMR metabolomics QC indicators      249 fields  max n 412,098    345 extract columns
  cat   222  NMR metabolomics processing          12 fields  max n 488,521     21 extract columns
  cat  1839  Protein biomarkers                    5 fields  max n  53,039     20 extract columns

Five files on disk: the three raw schema tables, the filtered CSV (517 rows, 187,261 bytes, 76 of them flagged n_is_sentinel), and manifest.json. Keep the raw TSVs — the catalogue is rebuilt continuously and the version column moves per field, so a checksummed copy is the only way to tell a real change from a mis-parse later.

The error path is worth exercising too. QUERY = r"metabolomic aging clock" exits 1 with no field matched …, because titles name measurements, not constructs — a clock, a risk score you intend to derive, or a phenotype definition has no field to match. Search for the inputs instead. Conversely QUERY = r"c-reactive protein" returns 7 fields across Blood biochemistry (n up to 469,326) and its processing category, so plenty of names you would assume are absent are simply spelled UK Biobank's way.

The query above is a flattering one: all 517 of its fields are availability = full, so it never exercises the retired path. Run it once on something that does, and read the summary rather than the CSV:

86 fields matched /hospital inpatient|episodes containing/
  cat  2000  Hospital inpatient                   36 fields  max n 413,163     36 extract columns  36 RETIRED
  cat  2002  Summary Diagnoses                    10 fields  max n 409,809     10 extract columns  10 RETIRED
  cat  2003  Summary Maternity                    23 fields  max n 195,756     23 extract columns  23 RETIRED
  cat  2004  Summary Psychiatric                   6 fields  max n 365,945      6 extract columns  6 RETIRED
  cat  2005  Summary Operations                   11 fields  max n 409,747     11 extract columns  11 RETIRED

Eighty-six for eighty-six, at cohort-scale participant counts, and every one of those categories shows an empty field list on the Showcase website. Without the RETIRED count this reads as the richest result in the skill.

Requesting access

This skill cannot obtain access to UK Biobank data, and nothing in it should be read as a promise of access. What follows describes the published route so you can decide whether to start it. The authoritative documents are the Access Procedures PDF and the UK Biobank website, linked under Sources; where this section and those disagree, they are right and this is stale.

Applications are closed. Do not start this process expecting it to complete. Two UK Biobank pages said so on 2026-08-18, and they are the two a reader would land on:

Access Management System login page: "We are currently not accepting new applications to UK Biobank, as the UKB-RAP remains closed. We will share an update via the Researcher Community as soon as new applications reopen."
Apply for access (page's own stamp 1 July 2026): "Applications are currently paused whilst necessary changes are made to the UK Biobank Research Analysis Platform, and our priority is restoring access to compliant researchers on ongoing projects." followed by "We intend to accept new applications in late 2026."

So there is a published intention but no open door and no committed date. Everything below describes the route as published, for deciding whether to queue for it — re-read both pages before planning a timeline. This is a dated observation of a state expected to change, and the catalogue work above is the part you can do today regardless.

Who may apply. Any bona fide researcher, for health-related research in the public interest, from academia, charity, government or commercial industry, in any country, all under the same process and criteria. The legal counterparty is the Applicant Institution, not the individual; the Applicant PI is named but carries no direct contractual responsibility. Collaborating institutions can be added at any time, each registering its researchers, executing the agreement, and paying an additional institution fee.

The published eight steps. Register on AMS → registration reviewed (the Access Team aims for 5 days) → complete the application form and select the data → application reviewed by the Access Team and the Scientific Team → approved → fees and material transfer agreement issued → payment and signatures returned → data released. Approval is valid for 90 days, within which the fee must be paid and the agreement signed.

What the forms ask for. Registration: name, address, email, a CV, a list of peer-reviewed publications with links where possible, and the research department and institutional website. The application: a short lay summary; every researcher and institution involved; a summary scientific rationale — research question and aims, background and any pilot data, a brief overview of planned methodology, and the expected value and public health impact; contact details for a signatory authorised to sign the agreement for each institution; and the selection of the data itself, which is where the catalogue work above pays off.

Two things this list does not include, and it is worth knowing early: your own ethics approval, and a hypothesis review. UK Biobank's Research Tissue Bank registration covers data applications, and the Access Procedures state UK Biobank does not consider its remit to be second-guessing the science "except in situations where the application is potentially untenable, absurd or unethical".

Duration and ongoing obligations. Projects are granted a minimum three years, extendable in one-year increments during the final year and cumulatively beyond that, subject to fees and compliance. Annual report forms are required. Findings must be published in a journal or an open-access site — the Access Procedures ask for reasonable endeavours to do so within 6 months of the project completion date, and a commensurate level of publication within the first three years — and results must be returned to UK Biobank so other researchers can use them. Researchers agree not to attempt to identify participants. UK Biobank does not approve publications, and makes no claim over inventions developed using the resource.

Fees. Cost-recovery only; UK Biobank states it recovers the incremental cost of servicing an application, not the cost of building the resource. The structure is three tiers by dataset size, charged for an initial 3-year period and renewable, exclusive of VAT, plus a per-additional-institution fee. Sample and re-contact applications are priced case by case. The tier amounts are on the fees page, not in the Access Procedures PDF. As that page carried them on 2026-08-18 (its own stamp: 14 April 2026):

	first 3 years	per year extension	covers
Tier 1	£3,000	£1,000	questionnaires, physical measures, health outcomes, linked health data
Tier 2	£6,000	£2,000	plus assays, proteomics, measured and imputed genotypes
Tier 3	£9,000	£3,000	plus imaging, large-scale assays, whole genome and exome sequence
additional institution	£1,000	£500	each institution added to an application
student / lower-income country	£500	£175	reduced access fee

Two things that page says which are easy to get wrong from an older reading. All tiers now include access via the UKB-RAP only — that is not a restriction peculiar to the reduced-fee route; downloading fields at all is an exceptional-circumstances request charged at the Tier 3 fee. And the tier boundaries are currently discounted: proteomics sits under Tier 1 and imaging under Tier 2 "for a limited period", so a field's cost_on value and the tier you are billed may not be the same thing this year. Read the fees page rather than trusting any figure quoted anywhere else, this table included; the cost_on column tells you which tier each field is classified in.

What this skill will and will not do for you. It will assemble your field shortlist with participant counts and fee tiers, tell you which fields are tabular versus bulk versus portal-table, draft a research question and methodology summary for you to edit, and checklist what each form asks for. It will not fill in or draft any attestation — data security arrangements, institutional or ethics status, non-re-identification undertakings, or the signature of an authorised signatory. Those are legal claims made by a named person on behalf of an institution, and they have to be written by that person.

Try it

A self-contained check that the catalogue is reachable and still has the shape this skill describes. Open data, no account, no key.

Data — UK Biobank Showcase schema 1, the data-field catalogue:

https://biobank.ndph.ox.ac.uk/showcase/scdown.cgi?fmt=txt&id=1

Openly published field-level metadata for the whole resource — no participant data, no account, no application. The Showcase publishes wget commands for these schema files on its own schema pages, so retrieving them is what the endpoint is for. They do not carry an open-data licence: the Showcase footer reserves reuse without a written licence from UK Biobank, so treat the downloaded catalogue as reference material for your own planning and do not republish or redistribute it. Last confirmed reachable 2026-08-18, together with schema 3 (categories) from the same host.

import csv, io, re, urllib.request
from collections import Counter

SHOWCASE = "https://biobank.ndph.ox.ac.uk/showcase"

def schema(sid, encoding):
    """Showcase schema tables are TSV. Encoding differs per table — see below."""
    raw = urllib.request.urlopen(f"{SHOWCASE}/scdown.cgi?fmt=txt&id={sid}", timeout=180).read()
    # scdown.cgi answers HTTP 200 with an HTML error page for a table id it does not know.
    assert not raw.lstrip()[:9].lower().startswith(b"<!doctype"), f"schema {sid}: HTML, not TSV"
    text = raw.decode(encoding)
    # QUOTE_NONE: 8,756 literal double-quote characters live inside titles and notes.
    return raw, list(csv.DictReader(io.StringIO(text), delimiter="\t", quoting=csv.QUOTE_NONE))

raw_fields, fields = schema(1, "utf-8")      # data-field properties — valid UTF-8
raw_cats,   cats   = schema(3, "cp1252")     # categories — NOT UTF-8, cp1252

COLUMNS = ["field_id", "title", "availability", "stability", "private", "value_type",
           "base_type", "item_type", "strata", "instanced", "arrayed", "sexed", "units",
           "main_category", "encoding_id", "instance_id", "instance_min", "instance_max",
           "array_min", "array_max", "notes", "debut", "version", "num_participants",
           "item_count", "showcase_order", "cost_do", "cost_on", "cost_sc"]
assert list(fields[0].keys()) == COLUMNS, list(fields[0].keys())

cat_name = {c["category_id"]: c["title"] for c in cats}

print("bytes (field catalogue) :", len(raw_fields))
print("fields                  :", len(fields))
print("categories              :", len(cats))
print("catalogue version (max) :", max(f["version"] for f in fields)[:10])

# num_participants is a SENTINEL field, not a plain count.
n = [int(f["num_participants"]) for f in fields]
print("num_participants == -1  :", sum(v == -1 for v in n), "(sentinel, not a count)")
print("num_participants ==  0  :", sum(v == 0 for v in n))
print("largest num_participants:", max(n))

# main_category is an integer id, and not every id used by a field exists in schema 3.
orphan = sorted({f["main_category"] for f in fields} - set(cat_name))
print("main_category ids used but absent from schema 3:", orphan)

print("instanced:", dict(sorted(Counter(f["instanced"] for f in fields).items())),
      " arrayed:", dict(sorted(Counter(f["arrayed"] for f in fields).items())))

# The orphan category is half LINKED and half UNLINKED — the distinction that decides
# whether an imaging file can be joined to a participant at all.
orphans = [f for f in fields if f["main_category"] == "1000"]
print("category 1000 titles    :",
      dict(Counter(f["title"].split(":")[0].strip() for f in orphans)))
assert len(orphans) == 126 and len(set(Counter(
    f["title"].split(":")[0].strip() for f in orphans).values())) == 1

# RETIRED FIELDS ARE IN THE DOWNLOAD AND NOT ON THE WEBSITE. A search that ignores
# `availability` silently offers dead field ids, sometimes under the live field's own title.
retired = [f for f in fields if f["availability"] == "4"]
live_titles = {f["title"] for f in fields if f["availability"] != "4"}
shadowed = sorted((f["field_id"] for f in retired if f["title"] in live_titles), key=int)
print("availability            :", dict(sorted(Counter(f["availability"] for f in fields).items())))
print("retired fields          :", len(retired))
print("retired under a live field's exact title:", len(shadowed), shadowed[:4], "...")
assert shadowed, "no superseded titles — re-check availability before trusting a shortlist"
assert "189" in shadowed                       # 189 and 22189 are both 'Townsend deprivation
assert "22189" not in shadowed                 # index at recruitment', both n=501,315

# Ground truth for that: the Showcase's own category page lists live fields only.
page = urllib.request.urlopen(f"{SHOWCASE}/label.cgi?id=126", timeout=180).read().decode(
    "utf-8", "replace")
# Parse the LISTING TABLE, not every field.cgi link. A category's prose cross-references
# other fields, and on cat 150 it names two retired ones in a sentence about retirement.
listed = set(re.findall(r'<tr class="row_(?:odd|even)"[^>]*id="f(\d+)"', page))
in_cat = [f for f in fields if f["main_category"] == "126"]
dead = [f for f in in_cat if f["availability"] == "4"]
print(f"category 126 Liver MRI  : schema 1 has {len(in_cat)}, its Showcase page lists "
      f"{len(listed)}, difference {len(dead)} retired")
assert len(in_cat) - len(dead) == len(listed), (len(in_cat), len(dead), len(listed))

# 126 passes with either method, so it cannot detect the method rotting. 149 and 150 can:
# a `field.cgi?id=` scrape returns 29 and 7 there against a true 28 and 5.
for cat in ("149", "150"):
    p = urllib.request.urlopen(f"{SHOWCASE}/label.cgi?id={cat}", timeout=180).read().decode(
        "utf-8", "replace")
    tbl = set(re.findall(r'<tr class="row_(?:odd|even)"[^>]*id="f(\d+)"', p))
    live = [f for f in fields if f["main_category"] == cat and f["availability"] != "4"]
    scraped = set(re.findall(r"field\.cgi\?id=(\d+)", p))
    print(f"category {cat:<3}             : table {len(tbl)} == live {len(live)}, "
          f"prose scrape would say {len(scraped)}")
    assert len(tbl) == len(live), (cat, len(tbl), len(live))
    assert len(scraped) > len(live), f"cat {cat} no longer distinguishes the two methods"

# scdown.cgi is an HTTP-200-with-an-error-body endpoint. The status code proves nothing.
bad = urllib.request.urlopen(f"{SHOWCASE}/scdown.cgi?fmt=txt&id=9999", timeout=60)
body = bad.read()
header = csv.DictReader(io.StringIO(body.decode()), delimiter="\t").fieldnames
print(f"scdown.cgi?id=9999      : HTTP {bad.status}, {bad.headers.get('Content-Type')}, "
      f"parses as {len(header)} column named {header[0]!r}")
assert bad.status == 200 and body.lstrip()[:9].lower().startswith(b"<!doctype")

# The two deliverables: where NMR metabolomics and Olink proteomics actually live.
for probe in ("NMR metabolomics", "Protein biomarkers"):
    ids = [cid for cid, t in cat_name.items() if t == probe]
    hits = [f for f in fields if f["main_category"] in ids]
    best = max(hits, key=lambda f: int(f["num_participants"]))
    print(f"{probe:20} category {ids} | {len(hits):4} fields | "
          f"max n = {int(best['num_participants']):,} (field {best['field_id']})")

Expect

Invariants — these hold across rebuilds, and a failure means this skill is wrong:

Schema 1 parses as TSV with exactly the 29 columns in COLUMNS, in that order. The assertion is the test; the documentation still lists a 28th column named tier.
Schema 1 decodes as UTF-8 and schema 3 does not — swap the two encodings and the cp1252 line raises UnicodeDecodeError. This is the trap the block exists to pin.
num_participants contains negative values, so it is a sentinel column. Its maximum is above 400,000, because some field covers nearly the whole cohort — a maximum well below that means the file is truncated or the columns are mis-aligned.
orphan is non-empty: at least one main_category used by a field is absent from schema 3, so a category join must tolerate misses.
instanced takes values 0, 1, 2 and arrayed takes 0, 1 — never anything else, and they vary independently.
Category 1000 splits evenly between EMBARGOED UNLINKED and EMBARGOED LINKED. The assertion is the test. An earlier version of this skill said all 126 were unlinked, which reversed the linkage claim for 63 imaging fields.
shadowed is non-empty, contains 189, and does not contain 22189. Retired fields sit in the download under the live field's own title, and 189/22189 are the pair that proves it — same title, same n=501,315, one of them dead. If this list ever empties, do not assume the problem went away; check availability before believing it.
Category 126's field count in schema 1 minus its retired fields equals what label.cgi?id=126 lists. 11 − 4 = 7. This is the ground truth that the Showcase website hides retired fields and the download does not, and it holds on all 22 categories that contain one.
scdown.cgi?fmt=txt&id=9999 returns HTTP 200 with HTML, and a TSV parser turns it into a one-column table named <!DOCTYPE HTML> without raising. Status codes are not a validity test on this host, for the schema download any more than for field.cgi.
NMR metabolomics and Protein biomarkers each resolve to exactly one category id, and Protein biomarkers holds fewer than 10 fields — the Olink values are a portal record table, not per-protein fields.

Observed 2026-08-18 against catalogue version 2025-09-20 — these move whenever UK Biobank rebuilds, so a mismatch is drift to investigate, not a failure:

bytes (field catalogue) : 4164390
fields                  : 11821
categories              : 410
catalogue version (max) : 2025-09-20
num_participants == -1  : 270 (sentinel, not a count)
num_participants ==  0  : 12
largest num_participants: 501938
main_category ids used but absent from schema 3: ['1000']
instanced: {'0': 4605, '1': 7198, '2': 18}  arrayed: {'0': 11271, '1': 550}
category 1000 titles    : {'EMBARGOED UNLINKED': 63, 'EMBARGOED LINKED': 63}
availability            : {'0': 11592, '1': 54, '4': 172, '7': 3}
retired fields          : 172
retired under a live field's exact title: 11 ['189', '20033', '20034', '20074'] ...
category 126 Liver MRI  : schema 1 has 11, its Showcase page lists 7, difference 4 retired
category 149             : table 28 == live 28, prose scrape would say 29
category 150             : table 5 == live 5, prose scrape would say 7
scdown.cgi?id=9999      : HTTP 200, text/html, parses as 1 column named '<!DOCTYPE HTML>'
NMR metabolomics     category ['220'] |  251 fields | max n = 488,514 (field 23480)
Protein biomarkers   category ['1839'] |    5 fields | max n = 53,039 (field 30900)

Sources

Showcase — https://biobank.ndph.ox.ac.uk/showcase/
Schema index (every downloadable table and its column definitions) — https://biobank.ndph.ox.ac.uk/showcase/schema.cgi
Schema 1 column definitions and code meanings — https://biobank.ndph.ox.ac.uk/showcase/schema.cgi?id=1
Mirror — https://biobank.ctsu.ox.ac.uk/crystal/
Access Procedures v2.1 (July 2022) — https://www.ukbiobank.ac.uk/wp-content/uploads/2026/05/Access-procedures.pdf
Apply for access — https://www.ukbiobank.ac.uk/use-our-data/apply-for-access/
Fees — https://www.ukbiobank.ac.uk/use-our-data/fees/
Financial support — https://www.ukbiobank.ac.uk/use-our-data/fees/financial-support/
Access Management System — https://ams.ukbiobank.ac.uk/ams/
Researcher community — https://community.ukbiobank.ac.uk/hc/en-gb
Sudlow et al. (2015) PLoS Medicine 12, e1001779 — https://doi.org/10.1371/journal.pmed.1001779
Bycroft et al. (2018) Nature 562, 203-209 — https://doi.org/10.1038/s41586-018-0579-z
Littlejohns et al. (2020) Nature Communications 11, 2624 — https://doi.org/10.1038/s41467-020-15948-9
Backman et al. (2021) Nature 599, 628-634 — https://doi.org/10.1038/s41586-021-04103-z
Julkunen et al. (2023) Nature Communications 14, 604 — https://doi.org/10.1038/s41467-023-36231-7
Sun et al. (2023) Nature 622, 329-338 — https://doi.org/10.1038/s41586-023-06592-6
Oh et al. (2023) Nature 624, 164-172 — https://doi.org/10.1038/s41586-023-06802-1

The field catalogue is published openly by UK Biobank and needs no account, but it is not released under an open-data licence — the Showcase carries a legal notice reserving reuse without a written licence. Download it, plan with it, and do not redistribute it. Individual-level participant data is separate, controlled, and governed by the material transfer agreement issued with an approved application. Cite the resource papers above and your application reference in published work.

uk-biobank

UK Biobank field catalogue

Read this before you plan a study

Downloading the catalogue

The 29 columns, and the code tables

Searching and grouping — the triage layer

Retired fields are in the download and not on the website

instanced and arrayed — the part that trips people

Resolving a field to its showcase page

Traps confirmed on the live file

Two worked answers

Get the files

Requesting access

Try it

Sources

`instanced` and `arrayed` — the part that trips people