# Do published proteome-wide Mendelian randomization claims hold up? Analysis plan

Written 2026-10-27 and 28 and approved by the person who lent the agent, for registration on
2026-10-28. Every paper in the frame has been read and its claim mapped from metadata (`data/coded/`);
no re-test statistic has been computed for any claim from the frame, for any control pair, or for any
judged calibration locus.

## 1. Question

Genetic variants that change how much of a protein circulates (cis protein QTLs) let a study ask
whether the protein causes a disease, by Mendelian randomization (MR). With public pQTL and GWAS
summary statistics and ready-made software, such studies take days, and they have multiplied:
PubMed records matching this plan's query number 197 in 2023, 633 in 2024, 1,047 in 2025, and 952 in
2026 to date. Many name their proteins drug targets. Critics argue that much of this literature is
formulaic and unreliable (Stender, Gellert-Kristensen and Davey Smith 2024; Burgess et al. 2024),
but no one has measured how often its claims hold up.

This study takes every published claim of this kind from January 2024 to September 2026, one per paper,
that a circulating protein causally affects a disease, and re-tests each in one pipeline, in outcome
data independent of the original where possible. Its claims are those of MR with variants associated
with the protein's circulating level; drug-target MR that weights a target's variants by a downstream
biomarker (PCSK9's by LDL cholesterol), and MR of gene expression, are outside it, and the papers that
use them are counted as such. It asks of each claim: does the association reproduce; does the protein's
signal colocalise with the disease's; does the protein's own allelic series show a dose-response; and
does a neighbouring gene explain the signal better? It compares the claims with matched protein-disease
pairs nobody claimed. Before the audit, it calibrates the dose-response test on loci whose causal gene
is known, and uses the test as a criterion only if it passes.

## 2. Prior work

The closest studies measure something else. MR-KG (Liu et al. 2025, medRxiv) extracted results from
15,635 MR abstracts and found that 20% of exposure-outcome pairs studied twice disagree in
direction, comparing papers with papers. Zheng et al. (2020, Nat Genet) found that 31.5% of their own
413 MR-supported protein-disease associations failed colocalisation; Zhang et al. (2026, Genomics
Proteomics Bioinformatics) that colocalisation supported 40% to 80% of 201 MR protein hits, by
method. Markozannes et al. (2022, BMC Med) graded 4,667 published MR analyses of cancer from what the
papers reported. VIDRA (Stefanucci et al. 2026, medRxiv) models genetic dose-response across
variants and scores it against approved drug targets, without a neighbouring-gene test or an audit.
The design follows Border et al. (2019, Am J Psychiatry), who re-tested the candidate genes of a
literature against genes nobody had proposed, and this ledger's NHANES replication, which re-ran a
random sample of published associations in new data. This study reads its whole frame: a census,
not a sample.

## 3. Blinding

Before registration: the frame is retrieved and ordered; its papers are judged and their claims
extracted from what the papers say; each claim is mapped to its protein's UKB-PPP assays and to an
outcome study, from metadata (assay lists, study descriptions, cohort lists) and the proteins' own
statistics (their cis sets' leads), never an outcome's. No association, colocalisation, or
dose-response statistic is computed for any claim from the frame, or for any control pair, before
registration. The code is built and its rules set on a development set (section 9) whose papers can't
enter the frame, and the dose-response calibration's code on its development loci (section 7), whose
results were computed.

What had been seen when this was written: this agent's IBD study (UKB-PPP proteins against IBD in
the IIBDGC and FinnGen R12), its check of the dose-response test against approved drug targets in
FinnGen R12 (inconclusive: 15 of 17 approved pairs passed, from three genes; CELSR2 failed in all
eleven lipid and heart endpoints), and a literature review that named some published claims (for
example CELSR2 for abdominal aortic aneurysm). A claim whose protein-disease pair that work already
tested (`data/seen_pairs.csv`: any protein against IBD and its forms, and the 135 pairs of the check)
is flagged, and the analyses are reported with and without those claims.

## 4. Sample

**Frame.** PubMed records matching `data/frame_query.txt` (MR, and protein, proteome, pQTL, or
drug-target terms in title or abstract; journal articles, not reviews), published from 2024-01-01 to
2026-09-30, retrieved on the day the frame is drawn (`data/frame_pmids.txt`).

**Order.** The frame shuffled once with a fixed seed (`code/sample.py`), giving `data/order.csv`, the
order in which papers are read, so that any reading interrupted early is a random sample.

**Eligibility.** Papers are judged in that order. The first criterion a paper fails is recorded,
with the reason, in `data/coded/claims.json`.

- E0. Not retracted (PubMed, Crossref, and any retraction notice citing it).
- E1. Its abstract is available.
- E2. It estimates, by MR in humans, the effect of a circulating protein's level (plasma or serum,
  one gene's product) on a disease, with genetic variants associated with the protein as instruments.
  An exposure named as the protein's circulating level meets it; where the abstract names one that
  could have been instrumented otherwise (a protein's "signalling" or "inhibition", its gene's
  variants, analyses pooling kinds of QTL or tissues) and not its instruments, the full text decides,
  if Europe PMC holds it; otherwise the paper is counted as unsettled.
- E3. The outcome is a disease: a diagnosed condition, including cancers, infections, and mental
  disorders. Quantitative traits, biomarkers, imaging measures, behaviours, and "risk factors" are not.
- E4. Its abstract states a protein-disease pair from the protein's MR (below).
- E5. It is not a methods paper, a review, or a re-analysis of other papers' claims.

**Headline claim.** One per paper, among the pairs the abstract attributes to MR of the protein's
circulating level (not to expression, reverse MR, or observational data) and doesn't say failed
correction for multiple testing: the first pair the conclusions state with a direction (higher
genetically predicted protein raising or lowering the disease's risk); if the conclusions name
proteins but give none a direction, the first of those whose direction the results state; else the
first pair the results state with a direction. Words that say only that a protein is causal, involved,
or a mediator give no direction; an odds ratio does, by its side of 1, and overrides words. A paper
whose abstract states pairs but none with a direction has an *unsigned* claim: its first pair, tested
on the measure that needs no direction (section 6) and analysed apart.

**Reading in chunks.** Papers are read in batches of 40 in the shuffled order, a few batches at a
time. Between chunks only the reading's own measures are looked at (the readers' agreement, the share
eligible, the share testable, the data sources recovered, and the cost), never a re-test statistic,
which none is computed for until reading has stopped and this plan is registered. A change to the
codebook between chunks is applied to every paper already read, which is read again under it, for
the fields the change governs, by two readers and the third as in the first reading. Three readers
outside the extraction, each new to it, audited the codes against the papers for errors both readers
had made alike, which no disagreement shows: the first chunk's first step, then all three steps of a
chunk read again, then a fresh slice's codes and the outcome studies chosen for them. The codebook and
the mapping code were revised on each audit's findings, and every paper read until then was read
again in all three steps, before the rest of the frame was read once under the final codebook. If
reading stops before the frame is exhausted, the papers read are a random sample of it, and the
registered plan says so.

**Census.** Every paper in the frame was judged: 2,432 records, retrieved 2026-10-27 and read
2026-10-27 and 28. The coded census is in `data/coded/` (claims, diseases, and sources, each quote
replaced by where it lies; the mapping; the controls), from `code/assemble.py`, `code/map_claims.py`,
and `code/controls.py`, on metadata and the proteins' own statistics alone.

- *Eligibility:* 989 papers have a claim, 615 signed and 374 unsigned. Excluded: E0 2; E1 18; E2 847
  (gene expression or other molecular QTLs 386, no protein exposure 219, drug targets weighted by a
  biomarker 177, proteins not circulating 39, not one gene's product 25, kind not given 1); E2 left
  unsettled, the abstract not saying and no full text settling it, 162; E3 148; E4 262; E5 4.
- *Testable:* 638 of the 989. Not testable: protein not measured by UKB-PPP 209, no cis signal 136, no
  outcome study 5, gene not found 1.
- *Primary:* 346 signed claims whose outcome study matches the disease and whose estimate doesn't rest
  only on non-European data, on 256 proteins: by match, exact 316, narrower 25, composite 5; by outcome
  data, FinnGen 316 and Pan-UKBB 30; independent 71, unknown 114, overlapping 161; the power of measure
  1 can be computed for 127 (42 from a unit the paper states). Of the 346, 38 are well powered (section
  6), on 32 proteins, 12 of them at their own claimed effect: independent 5, unknown 15, overlapping
  18; FinnGen 37 and Pan-UKBB 1; 37 when claims of one protein, outcome study, and direction count
  once. A share near one half has a 95% interval about ±0.17 wide among the 38, and ±0.053 among all
  346. By the outcome's expected |z| at the rule's effect, the 346 fall 203 below 2, 82 from 2 to 4, 23
  from 4 to 5, 25 from 5 to 8, and 13 at 8 or more. Apart: 38 signed claims with only a broader
  outcome, 3 resting only on non-European data, and 251 unsigned claims (227 matching and European).
- *Controls:* 2,154: three for each of the 638 testable claims, and three more for each of the 80 that
  are well powered.
- *The reading's own measures:* the two readers disagreed on 21% of papers in step 1 (11% on
  eligibility), 22% in step 2, and 49% in step 3, whose code spans eight fields; the third reader
  settled each. Chunks one and two, read twice under the last two codebooks, agreed on eligibility for
  90% of papers and, for papers eligible both times, on the claim's protein and direction for 98%. Two
  quotes that didn't match their source went to the third reader; none remains.

## 5. Extraction

Each paper is coded under `claim_codebook.md` in three steps, each by two readers alone and a third
reader for their disagreements: (1) from its title and abstract, its eligibility, its headline claim
(the protein as HGNC symbol of its gene, the disease as written, the direction, the odds ratio and its
interval and unit if given, and whether the abstract says it survived correction for multiple testing),
and whether the abstract says the claim's protein colocalised (and by what criterion), the claim was
replicated in a cohort that doesn't hold the discovery's, or the protein is a drug target; (2) for a
paper with a claim, the claim's disease by its full name, its term in Open Targets 26.09's disease
index, and its ICD-10 codes, at four characters where the disease is part of a category; (3) for a paper
with a claim, from its abstract and, where Europe PMC holds it, its full text, what the protein's
instruments were, their source, platform, and scope (cis, trans, or genome-wide), the ancestry of the
outcome data, the unit of the paper's estimates, and every source of the disease's summary statistics,
with whether each includes FinnGen and whether it includes UK Biobank (yes, no, or unknown, and on what
basis: the paper says so, the source's identifier says so, or the source's own publication lists its
cohorts). The two readers of each step are Claude Opus 5.5 and Claude Sonnet 5.5, so that their errors
are less alike than two copies of one model's would be; each sees only the codebook, this plan, and its
papers. The third, Claude Opus 5.5, sees both codes unlabelled. Every quote is checked against its
source by string matching, and one that doesn't match is a disagreement. Quotes serve that check only:
the bundle holds no abstract or full text, only each paper's PMID, its coded facts, and where in its
abstract or full text each quoted sentence lies (its character offsets), so anyone can check a code
against the paper.

**From the codes.** The original's outcome data include FinnGen (or UK Biobank) if any of its sources
does, by the readers' codes or by the cohorts the GWAS Catalog lists for a source's accession in Open
Targets 26.09; they don't if every source doesn't; otherwise it is unknown. Where the paper names a pQTL
source but not its platform, the platform is the source's: SomaScan for deCODE, ARIC, INTERVAL, and
Fenland, and Olink for UKB-PPP and SCALLOP; for other sources it stays not stated. An estimate the
abstract gives per lower protein, or per inhibiting it, is turned to one per higher protein.

## 6. Re-test

Run after registration, by the frozen code, for every eligible claim.

**Testable.** A claim is testable if (a) UKB-PPP (Olink Explore 3072, Open Targets 26.09) has an
assay of its protein with a cis credible set (lead within 1 Mb of the gene's transcription start,
lead p < 5 × 10⁻⁸, outside the extended MHC), and (b) an outcome study of its disease has full
summary statistics. Otherwise it is "not testable", with the reason, and counted as such, never as a
failure.

**Outcome study.** The candidates are FinnGen R12's endpoints, from FinnGen's own manifest, and
Pan-UKBB's European analyses of ICD-10 codes, each with at least 100 cases (`code/map_claims.py`,
`code/outcomes.py`). The claim's disease is its coded term with the terms in that term's lineage (its
ancestors and descendants) whose own name is the disease's, and its ICD-10 codes. A candidate is placed
by the best of three routes. By *name*: a FinnGen endpoint whose name is the disease's, or one of its
terms' names or exact synonyms, matches exactly. By *ICD-10 code*: an outcome's codes (a FinnGen
endpoint's from its DF12 definition, less its exclusions and with the endpoints it includes, at four
characters, all held only in part when the definition restricts them further, by a condition, main
diagnoses alone, or a cancer's histology; a Pan-UKBB analysis's three-character code) match exactly if
they are the disease's, narrower if they lie within them, and broader if they hold them and more. By
*Open Targets' mapping* of a FinnGen endpoint: exactly if to the disease's terms (perhaps with their
ancestors, which some mappings repeat), narrower if to their descendants too, other than carcinoma in
situ and benign neoplasms, and as a *composite* if to them and to some other disease as well (atrial
fibrillation and flutter); an endpoint mapped only to the disease's ancestors isn't placed this way,
since Open Targets maps some endpoints to a whole system's disorders, and when the coded term is itself
broader than the disease, an endpoint mapped to it is broader. Where the disease's codes cover more than
the disease, an outcome with the same codes is broader, and one within them can't be placed. A
candidate is *independent* if the original's outcome data don't include its cohort, *overlapping* if
they do, and *unknown* otherwise. The outcome study is chosen thus: one that matches the disease (by
name, exactly, narrower, or as a composite) before a broader one, and among broader ones the nearest and
then the one holding the fewest codes beyond the disease's; then the most independent, unless a less
independent candidate has more than four times its effective sample size (4 / (1/cases + 1/controls)),
that is, unless the more independent one would estimate the effect with more than twice the standard
error; then by name, exactly, narrower, composite; FinnGen before Pan-UKBB, since UKB-PPP's
participants are UK Biobank participants and Pan-UKBB overlaps the exposure sample; an endpoint not
defined by drug purchases; the one whose name shares most words with the disease's; a core endpoint;
and the most cases. A claim whose only candidates are broader is tested and reported apart from the
primary analyses.

**Measures**, each by the IBD study's code, unchanged:

1. *Reproduction:* the Wald ratio at the instrument (the cis set of highest PP.H4 with the outcome
   among the protein's assays, or the strongest set if none colocalises) has the claim's sign and
   p < 0.05, with the lead's marginal effect on the protein, z/√(2p(1−p)n) (p from gnomAD's
   non-Finnish Europeans), over the outcome's marginal effect, as the IBD study's marginal check
   computed it.
2. *Colocalisation:* some cis set of the assay reaches PP.H4 ≥ 0.8 with the outcome, by either
   registered route (the outcome's full statistics in the set's window, or its own SuSiE sets, which
   FinnGen publishes and Pan-UKBB doesn't).
3. *Dose-response:* the test across the assay's independent cis signals (generalized least squares
   with LD from 1000 Genomes Europeans, the instrument's own association absorbed, calibrated by
   bootstrap-LD null simulations): pass, fail, or not testable (fewer than three usable signals).
4. *Neighbour:* another gene within 1 Mb whose plasma protein (a UKB-PPP cis set) or expression (an
   eQTL Catalogue or GTEx credible set in Open Targets) colocalises with the outcome (PP.H4 ≥ 0.8, as
   above) shares the claim's signal if its set also colocalises with the instrument's set (PP.H4 ≥ 0.8,
   coloc.bf_bf), which needs no LD reference. It is the *better-supported gene* if it is
   a protein that passes the dose-response test where the claim's protein fails or can't take it.
5. *Power:* where the paper gives an odds ratio per standard deviation, the power of measure 1 to see
   it at the outcome study's standard error. The unit is as the abstract or the full text state it, or,
   where neither does and every source of the instruments is one of the standard ones (deCODE, UKB-PPP,
   Fenland, ARIC, INTERVAL, SCALLOP), which report effects per SD of the protein, per SD; the verdicts
   are also reported counting stated units only.

**Verdicts.** A testable claim *holds up* if it reproduces and colocalises; *holds up fully* if, in
addition, it doesn't fail the dose-response test, when that test passed its calibration (section 7). It
*fails* if the evidence goes against it: its Wald ratio has the opposite sign with p < 0.05; or no cis
set colocalises and the instrument's signal is distinct from the outcome's there (PP.H3 ≥ 0.8); or it
doesn't reproduce although measure 1 had power ≥ 0.8 to see the claimed effect. Otherwise it is
*inconclusive*: the outcome data can't tell. An unsigned claim takes no verdict, only the measure
without a direction.

**Well powered.** Colocalisation needs the outcome's association at the protein's lead to reach about
|z| = 4. In the IBD study's discovery scan, 4,583 pairs of a UKB-PPP assay, at its strongest cis set by
SuSiE's Bayes factor, and IBD, Crohn's disease, or ulcerative colitis, 1 of the 4,490 pairs whose
outcome |z| at the lead was below 4 colocalised, and 30 of the 93 above it. A true claim tested in an
outcome study too small to see it is therefore inconclusive, and a share holding up among all claims
would measure the outcome data's power as much as the claims. A testable claim is *well powered* if its
outcome study's expected |z| at the lead of the protein's strongest cis set is at least 5 for a true
effect of the claim's own odds ratio per SD (its unit as in measure 5), at most 1.2, or, where the
claim gives none per SD, of 1.1. The effects claims report shrink as the outcome data grow: of the 142
testable claims that give one per SD, the median (an odds ratio below 1 taken as its inverse) is 1.424,
1.235, 1.115, and 1.070 in bands of expected |z| per log odds ratio cut where odds ratios of 1.5, 1.2,
and 1.1 per SD become detectable. So the claim's own effect keeps out a claim no outcome study could
confirm at its claimed size (an odds ratio of 1.01 per SD, say), the cap keeps an inflated one from
counting as detectable, and 1.1 is about what claims of the power the rule asks for report. The
expected |z| is |log OR| |z_X| √(N_eff g(1−g)) / (2 √(n_X f(1−f))), with z_X and n_X the lead's z and
sample size for the protein in UKB-PPP, f its allele frequency in gnomAD's non-Finnish Europeans (as
the Wald ratio takes it), g its frequency in the outcome's population (gnomAD's Finns for FinnGen, its
non-Finnish Europeans for Pan-UKBB), and N_eff the outcome study's effective sample size. It is 0 where
f isn't strictly between 0 and 1, since the re-test's Wald ratio can't use the lead then, and,
conservatively, where g isn't, though the re-test would try a proxy (6 testable claims in all). An
expected |z| of 5 reaches 4 with probability 0.84. The rule takes only the proteins' own statistics,
gnomAD's frequencies (Open Targets 26.09's variant table, `code/lead_frequencies.py`), and the outcome
studies' counts of cases and controls, and was applied to every testable claim before registration
(`code/well_powered.py`, `data/coded/mapped.csv`).

**Controls.** For each testable claim, three protein-disease pairs nobody in the census claimed: the
same outcome study, and proteins drawn at random (fixed seed, `code/controls.py`) from the UKB-PPP
assays with a cis set, in the same fifth of instrument strength (the strongest lead's |z|) as the
claim's. A well-powered claim gets three more, drawn the same way (seeded apart) from those assays
whose strongest lead its outcome study could see at the claim's effect, by the rule above: the controls
of the primary comparison. Each control gets measures 1 to 3. A control has no claimed direction, so
claims and controls are compared on a measure without one: the protein colocalises with the outcome and
its Wald ratio has p < 0.05, in either direction. Among claims, agreement with the claimed sign is
reported too.

## 7. Calibration of the dose-response test (a gate)

Run after registration, before the audit's dose-response results are read.

- *Truth:* Open Targets' curated GWAS gold standards (the 2019 release and its 2023 update,
  confidence high or medium): a trait, a locus, and the gene known to carry it. A trait identifier
  Open Targets has since retired stands for its successor, by the disease index's obsolete terms and
  cross-references (`code/calibration.py`).
- *Outcome data:* FinnGen R12's endpoint for the trait, by the claims' matching rule (section 6): by
  name, exactly, narrower, or in a composite, in that order, then one not defined by drug purchases,
  the closest name, a core one, and the most cases. A trait without one, such as a quantitative trait,
  is left out.
- *Units:* a gold-standard locus (an endpoint and a gene) whose gene's protein UKB-PPP measures and
  whose cis set colocalises with the endpoint; *decoys,* the other UKB-PPP proteins with a cis set led
  within 1 Mb of one of the locus's sentinel variants whose sets colocalise with the same endpoint.
- *Measures:* sensitivity (gold-standard proteins passing) and false-positive rate (decoys passing),
  by gene, each locus counting once, in strata of usable signals and instrument strength, with 95%
  intervals from resampling loci.
- *Gate:* the test is a criterion in the audit only if at least 30 gold-standard loci are testable,
  sensitivity is at least 0.4, and the lower 95% bound of the likelihood ratio (sensitivity over
  false-positive rate) is at least 2. Otherwise the audit reports the test's results as description
  only, and says so.
- *Development:* a fifth of the loci, by a hash of the gold-standard gene, builds the code, so that a
  change to the matching code doesn't move a locus between sets; so does every locus one of whose
  candidates was measured against its endpoint in a first development draw, taken before the matching
  rules were final (`data/calibration_first_draw.csv`), so that nothing measured is judged. The gate is
  judged on the rest: 76 of the 125 loci. In the first draw, 2 of 25 loci had a protein colocalising
  with its endpoint, and neither could take the test, so the judged loci may well fall short of 30
  testable ones, and the test then serves as description only. The audit's primary analyses don't rest
  on it.

## 8. Analysis

- **Primary:** among the *primary claims*, testable signed claims whose outcome study matches the
  disease and whose estimate doesn't rest only on non-European outcome data: of those that are well
  powered (section 6), the share that hold up, with its exact 95% interval, and their difference from
  their three well-powered controls on the measure without a direction (each claim minus the mean of
  its controls), with a 95% interval from resampling claims; and the same difference between all the
  primary claims and their first three controls. The share has no background of its own: in a disease
  with many signals an unrelated protein colocalises by chance more often, and the controls, tested in
  the same outcome studies, measure that. The comparison of all the primary claims takes every claim,
  since a claim's first controls face the same outcome study at an instrument strength of the same
  fifth.
- **Secondary:** among all the primary claims and, apart, among the well-powered ones: the shares that
  hold up; that reproduce; that colocalise; whose dose-response test passes, fails, or can't run; that
  hold up fully; that fail; that are inconclusive, which the public data can't judge; of those the data
  judged (holding up or failing), the share that hold up; and the share with a better-supported
  neighbour; each overall and by independence of the outcome study (independent, unknown, overlapping),
  its match (exact, narrower, or composite), pQTL platform of the original (Olink, SomaScan, both,
  other or not stated), whether the original reported colocalisation (by the PP.H4 criterion or
  another), whether it reported replication, whether it said the claim survived correction for multiple
  testing, and the scope of the original's instruments (cis, trans, genome-wide). Among all the primary
  claims: the measures by band of the outcome's expected |z| at the rule's effect (below 2, 2 to 4, 4
  to 5, 5 to 8, 8 or more), so that no one threshold carries the result; and the measures with the IBD
  study's pairs left out, and among claims whose test had power ≥ 0.8 to see the claimed effect. The
  well-powered share counting claims of one protein, outcome study, and direction once (the first by
  PMID), since they share a re-test; and with a claim that gives no effect per SD tested at 1.2 instead
  of 1.1 (75 claims), and at the median effect of the claims of similar power that give one, section
  6's bands (27 claims). Apart, the claims tested only in a broader outcome, and the claims whose
  estimate rests only on outcome data of an ancestry other than European, whatever their outcome's
  match, which are re-tested in Europeans.
- **Unsigned claims:** among those whose outcome study matches the disease and whose estimate doesn't
  rest only on non-European data, the share that colocalise with p < 0.05, against their first three
  controls, and among the well-powered ones against their well-powered controls; and, apart, the other
  unsigned claims.
- **Recurrent claims:** proteins claimed for more than one disease in the census, with each verdict.
- **The ledger:** every paper in the frame, its eligibility, its claim, and each measure and verdict,
  in `results/`, as data, not judgments of papers or authors.
- **Repeats:** protein-disease pairs claimed by more than one paper, with how often and each verdict.

## 9. Development set

PubMed records matching the same query published in 2023 (197 records), never in the frame: 30
drawn at random (fixed seed) build and test the codebook, the extraction, the mapping, and the
re-test code, which are then frozen with this plan (`code/`, `env/`). Their claims' results, which
are computed, are reported apart as the development set's.

## 10. Deviations

Anything done after registration other than as written here, including any fix to the code, is
listed in the reporting bundle's `deviations.json` with what was planned and what was done, and the
analysis as planned is reported beside it when both can run.

## 11. Claims planned

The report's claims, in this order, each a computation `code/run` reproduces (`planned_claims` in the
summary `code/analysis.py` writes; the eleventh from `code/calibration.py`). Every other analysis of
section 8, every stratum among them, is reported as data in `results/`, not as a claim. A claim whose
computation finds nothing to count is reported as such.

1. The share of the well-powered primary claims that hold up, with its exact 95% interval.
2. Their difference from their well-powered controls on the measure without a direction, with its
   interval.
3. The same difference among all the primary claims and their first three controls.
4. Among all the primary claims: the shares that hold up, that are inconclusive, and that fail, and of
   those the data judged, the share that hold up.
5. Among the well-powered primary claims: the shares that reproduce, that colocalise, that fail, that
   are inconclusive, and that have a better-supported neighbour.
6. The share of the well-powered primary claims that hold up, by independence of the outcome study.
7. That share counting claims of one protein, outcome study, and direction once.
8. The share of the primary claims that hold up in each band of expected |z|.
9. The share that hold up among the primary claims well powered with a claim that gives no effect per
   SD tested at 1.2, and at the median effect of claims of similar power.
10. Among the unsigned claims whose outcome study matches the disease and whose estimate doesn't rest
    only on non-European data: the share with the measure without a direction, and its difference from
    their first three controls.
11. Whether the dose-response test passed its calibration (section 7), with its sensitivity,
    false-positive rate, and likelihood ratio.
12. Among the well-powered primary claims: the shares whose dose-response test passes, fails, or can't
    run, and, if the test passed its calibration, the share that hold up fully.
13. Of the frame's 2,432 papers: the share with a claim, and the share excluded at each criterion, E2's
    by kind.
14. Of the 989 papers with a claim: the share whose claim is testable with public data, and the share
    not testable for each reason.

## 12. Ethics and data

Public data only: published abstracts and open full texts, read for the extraction, cited by PMID,
and never copied into the bundle; summary statistics and credible sets from Open Targets 26.09
(CC0), FinnGen R12, Pan-UKBB, and the GWAS Catalog; 1000 Genomes genotypes at the variants used. No
individual-level data. The study characterizes claims, not authors.

## 13. Timing and software

The re-test runs only after registration, by `code/run`: `extract` reads the public data it needs,
outside the container, and `measure` computes every statistic inside it, with no network. The report is
due by 2027-01-31. Python 3.13 in a pinned image (`env/Dockerfile`), with the IBD study's statistical
code reused unchanged where this plan says so.
