[
 {
  "from": "prereg:94a41c8f07650e3ca3afafcb90f08a30990828a189d93e3bab4b2611f2d800f4",
  "kind": "changed",
  "detail": "Where the statistics run. The registered run, plan/code/run's measure step, ran retest.py, calibration.py's measure and summary steps on the judged loci, and analysis.py in the image env/Dockerfile builds, from the folder its extract step wrote. code/run runs the same four commands, unchanged, in the verifier's container, which the reference harness builds from the same Dockerfile: for the census from data/main and for the development set from data/dev, each copied to a scratch folder first, since retest.py and calibration.py write their results beside what they read, and with the results copied to results/. The declared results come from code/run. In the first version, from the registered mapping, every value it declared was the same as the registered run's; this version's come from the corrected mapping (below)."
 },
 {
  "from": "prereg:94a41c8f07650e3ca3afafcb90f08a30990828a189d93e3bab4b2611f2d800f4",
  "kind": "unstated",
  "detail": "The plan did not fix Python's hash seed. code/run sets PYTHONHASHSEED=0, so the registered code's iterations over sets of strings run in one order on every run; the registered run, without it, gave the same declared values."
 },
 {
  "from": "prereg:94a41c8f07650e3ca3afafcb90f08a30990828a189d93e3bab4b2611f2d800f4",
  "kind": "unstated",
  "detail": "data/ holds what plan/code/run's extract step wrote, the extracted inputs only: for the census (data/main) from the registered run of 2026-10-28, with the coded census copied from plan/data/coded, except what the correction changed (below): the mapping and controls in data/main/coded and data/main/retest, and the re-test's inputs that code/reextract.py extracted again; and the UKB-PPP protein sets, their leads' marginal effects, and their leads' frequencies in gnomAD's Finns (data/proteins). The development set (data/dev) was extracted again after registration with the same step and its final mapping, since its well-powered controls were drawn after its first extraction; its codes are exported without quotes, as the census's are (plan/code/export.py). data/sources.csv, written after registration, lists every public file the extraction read; the first version's listed 10 of the 22 chromosomes' 1000 Genomes files that the genotypes were read from, and this version's lists all 22, with the two FinnGen endpoints the correction read for the first time."
 },
 {
  "from": "prereg:94a41c8f07650e3ca3afafcb90f08a30990828a189d93e3bab4b2611f2d800f4",
  "kind": "changed",
  "claims": [
   "C1",
   "C2",
   "C3",
   "C4",
   "C5",
   "C6",
   "C7",
   "C8",
   "C9",
   "C10",
   "C12"
  ],
  "detail": "The correction. The plan (section 6) matches an outcome by name when its name is the disease's or one of its terms' names or exact synonyms, and the registered code (plan/code/outcomes.py, finngen_matches) did so even where the coded term is broader than the disease, which the codes mark, so an endpoint named by the broader term's names matched exactly, though the plan's other two routes count an endpoint mapped to such a term, or holding such codes, as broader, and leave a claim whose only outcomes are broader out of the primary analyses. A challenge to C1 found one case among the well-powered claims: F11 and early-onset stroke, matched to FinnGen's STROKE, at every age of onset. code/outcomes.py now counts an endpoint named by a broader term's names as broader, at the distance the mapping route gives it, and only one named as the disease itself as a match by name. With it, map_claims.py and controls.py, otherwise unchanged, give the mapping and controls in data/main/coded: 12 testable claims become broader, 7 of them primary claims and 1 of those well powered; three take a new outcome study, two of CCL27 and major adverse cardiovascular events (from FinnGen's other diseases of the pericardium to its hard cardiovascular diseases) and an unsigned one of APOE and dementia (from dementia to Alzheimer's disease, narrower), so 9 of their controls move with them, the CCL27 claim now well powered draws 3, and the APOE claim no longer well powered drops 3. The development set's mapping is unchanged. Every declared result was computed again from this mapping by the registered code; the first version's are in the bundle it corrects, results/ and results/main."
 },
 {
  "from": "prereg:94a41c8f07650e3ca3afafcb90f08a30990828a189d93e3bab4b2611f2d800f4",
  "kind": "changed",
  "detail": "The census's re-test inputs for the corrected mapping (data/main/retest) were extracted by code/reextract.py rather than by plan/code/run's extract step whole. It runs that step's functions in extract_retest.py, unchanged, on the corrected mapping and controls: the neighbours whole, from local data, and of the outcomes and genotypes only what the registered extraction doesn't hold. Four outcome studies' sets changed, FinnGen's hard cardiovascular diseases, Alzheimer's disease, dementia, and other diseases of the pericardium, and only those were read again; the other 227 studies' windows and lookups were carried over, since a study's depend on the study and its sets alone; and no new genotype was needed. Run on the registered mapping, the script rebuilds the registered extraction: its windows, lookups, and genotypes byte for byte, and its neighbour tables and FinnGen's own credible sets row for row, in another order, from which the registered re-test gives byte-identical results. Dementia and other diseases of the pericardium, read again on 2026-11-05, gave the registered extraction's statistics on every set they kept."
 },
 {
  "from": "prereg:94a41c8f07650e3ca3afafcb90f08a30990828a189d93e3bab4b2611f2d800f4",
  "kind": "unstated",
  "detail": "code/collect.py, written after registration, gathers the declared results: results/R1.json to R14.json for the plan's 14 claims in order, S1.json for the registered secondary analyses the paper quotes, E1.json for the exploratory checks, and D1.json for the development set. It computes nothing new; it rounds shares to percentages and differences to percentage points, each to one decimal, and rewrites the values the registered code wrote into its summaries as NaN or infinity, which JSON doesn't allow, as null."
 },
 {
  "from": "prereg:94a41c8f07650e3ca3afafcb90f08a30990828a189d93e3bab4b2611f2d800f4",
  "kind": "unstated",
  "claims": [
   "C1",
   "C2",
   "C3"
  ],
  "detail": "Exploratory checks, not in the plan, written after the results were known (code/exploratory.py): how the failing primary claims fail; the opposite-sign failures by the original's platform, instrument scope, and independence; the comparisons with controls of C2 and C3 within each level of the outcome study's independence; and a re-reading, by the same model family, of where each coded direction comes from, for the 38 well-powered primary claims and the 24 opposite-sign failures of the first version (data/main/direction_check.csv, which holds no words of any abstract); this version's file keeps the 56 of those 58 that are still among the well-powered primary claims and the opposite-sign failures, leaving out F11 and early-onset stroke and IL22 and abnormal spermatozoa, whose outcomes the correction made broader. The re-reading questions four directions: three taken from a proposed drug's action (IL6R and IL2RA inhibitors, an IFNLR1 agonist) and one where the abstract's words and odds ratio disagree (RSPO3). Read the other way through the registered verdict, 11 of the 37 well-powered claims would hold up rather than 10. The registered verdicts and declared results are unchanged."
 },
 {
  "from": "prereg:94a41c8f07650e3ca3afafcb90f08a30990828a189d93e3bab4b2611f2d800f4",
  "kind": "changed",
  "claims": [
   "C11",
   "C12"
  ],
  "detail": "Section 12 of the plan says 'No individual-level data', beside '1000 Genomes genotypes at the variants used'. The dose-response test as the plan registered it (section 6, measure 3: linkage disequilibrium from 1000 Genomes Europeans, calibrated by bootstrap-LD null simulations) reads the genotypes of the 503 European samples of the public 1000 Genomes panel at its leads (data/main and data/dev, kg_EUR.parquet in each set's folder) and resamples those samples for the uncertainty of linkage disequilibrium, so those files are individual-level reference genotypes, which the 1000 Genomes Project released for open use. The protein and disease analyses use only summary statistics, and the work uses no private individual-level data. The method is as registered; the plan's sentence was inaccurate, and the plan is left as registered."
 }
]
