Oncofusion

start / graph / ana-the-pmid-side-is-cleaner-than-the-doi-side-and-one-real-defect-was-a-1995-paper-stamped-2026

Analysisstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.

Running the wrong-paper check over the 17,925 PMID citations - the larger half of the corpus and the half where both documented failures happened - gives 100.0% year agreement and 99.7% on authors, one genuine defect, and shows the DOI side's 92% was an artefact of comparing against a single date field

also called the PMID side is clean · 100.0% year, 99.7% authors over 17,925 blocks · a 1995 paper stamped 2026 · Crossref issued is not PubMed pubdate · epubdate is a real publication year

The checks built over the last few cycles all ran on one kind of paper identifier. The other kind is more common here, and is the kind behind both mistakes this project has actually made. Running the same checks over it found essentially nothing wrong — one citation that had stamped this year onto a paper from 1995 — and revealed that an earlier result looked worse than it was, because the two databases mean different things by 'the year'.

The same thing, for a scientist

**Statement.** Cycles 123–126 bounded the *right identifier, wrong paper* failure over **DOIs**. **Both documented wrong-paper failures in this project were PMIDs**, and the corpus carries **more** PMID citations than DOI ones. The axis had never been run over them; the cycle-126 entry said so in its own open questions. | | PMID side | DOI side | |---|---:|---:| | citation blocks | **17,925** | 15,360 | | distinct identifiers | 3,663 | 3,438 | | **year exact** | **17,748 / 17,755 — 100.0%** | 15,104 / 16,420 — 92.0% | | year off by one | **7** | 1,294 | | **year disagreeing by 2+** | **0** | 22 | | **authors agree** | **16,232 / 16,281 — 99.7%** | 15,159 / 15,219 — 99.6% | **`P1` fires. The PMID side is at least as clean, and on year it is dramatically cleaner — which turns out to be a fact about the registries, not about the citations.** 🔵 **The DOI side's 92% was an artefact of comparing against one date field.** Crossref's `issued` is frequently the **online-first** date while the entry recorded the **issue** year — that is the whole of its 1,294 off-by-one. PubMed returns **both** `pubdate` and `epubdate`, and accepting either takes the PMID year rate from 99.1% to **100.0%**. **A citation is not wrong because two registries disagree about when a paper was published.** **One genuine defect, found and fixed.** `line-u937-calm-af10-adult-monocytic` carried a citation block with `pmid: 7759961` and `year: 2026` — the paper is **Reid YA, *J Leukoc Biol*, 1995**, a 31-year gap. **It is not a typo but a type confusion:** one block describing *two objects*, the Cellosaurus record accessed in 2026 and the 1995 paper Cellosaurus cites. Corrected to the paper's year and authors, with the Cellosaurus provenance moved into the title. 🟠 **All 49 author disagreements are placeholder strings**, the same class as the DOI side: *"Multicentre retrospective consortium"*, *"Project EVOLVE consortium"*, *"See PubMed record PMID 38989615 for the full author list"*. **None is a wrong-paper citation.**

Go deeper
Two registries, two meanings of 'the year'

PMID 34693799 was reported as a defect — stored 2021, resolved 2023. It is not:

pubdate  : 2023 Jan        <- journal ISSUE date
epubdate : 2021 Oct 23    <- ONLINE-FIRST

The entry's 2021 is correct. My checker compared pubdate alone. Accepting either date removed the disagreement and took the rate to 100.0%.

🔵 The same fix should be applied to the DOI check, where Crossref's single issued field produced 1,294 off-by-one blocks and 18 of the 22 larger gaps — all of which were read individually in cycle 123 and found to be legitimate online-versus-print splits. That reading was correct and the instrument was never adjusted. Doing so would raise the DOI year rate substantially and, more importantly, stop a standing checker crying wolf on ~1,300 citations every run.

🔴 The methodological point is general and this project keeps relearning it: a comparison between two sources is only a test of the data if both sources mean the same thing by the field. Cycle 119 found this for title (a descriptive sentence versus a paper title); here it is year.

What the placeholder class actually is, now that it appears on both axes

49 blocks on this side and 55 on the DOI side store something that is not an author list:

formexample
a consortium nameProject EVOLVE consortium
an explicit deferralSee PubMed record PMID 38989615 for the full author list
a verification noteNot verified beyond the PubMed record retrieved 2026-09-…
an internal pointerAs cited in protein-nup98-domains and localization-nucle…

🟠 These are not formatting noise and the previous entry was too quick to call them benign. A block whose authors field says see the PubMed record is a block where nobody transcribed the author list, which means nobody compared the record to what the entry claims it says. They are the subset where a wrong identifier would survive longest, and they are now identified twice over — 104 blocks across both axes — and still unread.

🔵 They are also the only population where these checks have any power at all. Everywhere else the stored value came from the registry being compared against, so agreement is near-tautological (U1). The instrument is strongest precisely where it currently abstains.

Facts and figures
questiondo the stored year and authors match the paper each PMID points at?
verdictP1 - the PMID side is at least as clean as the DOI side
blocks17925
distinct pmids3663
absence classmeasured-not-curated
bite backA 100.0% year rate was reached by accepting two dates as correct and fixing the one block that disagreed. Those two moves guarantee the number. The honest claim is that no PMID citation disagrees with a year PubMed itself admits - not that the years are right.
year
{
  "comparable": 17755,
  "coverage_pct": 99.1,
  "exact": 17748,
  "exact_pct": 100,
  "off_by_one": 7,
  "disagree_2plus": 0,
  "note": "exact accepts either the issue year or the online-first year"
}
authors
{
  "comparable": 16281,
  "coverage_pct": 90.8,
  "agree": 16232,
  "agree_pct": 99.7,
  "share_none": 49,
  "all_of_which_are": "placeholder strings, not author lists"
}
comparison to doi side
{
  "doi_year_exact_pct": 92,
  "doi_year_off_by_one": 1294,
  "why": "Crossref's `issued` is frequently the online-first date while the entry recorded the issue year; PubMed returns both pubdate and epubdate and accepting either removes the discrepancy - the DOI side's 92% is an artefact of a single date field"
}
the one real defect
{
  "entry": "line-u937-calm-af10-adult-monocytic",
  "pmid": "7759961",
  "stored_year": 2026,
  "true_year": 1995,
  "paper": "Reid YA et al., Cell line cross-contamination of U-937, J Leukoc Biol 1995",
  "diagnosis": "a type confusion, not a typo: one citation block describing TWO objects - the Cellosaurus record accessed in 2026 and the 1995 paper it cites",
  "fixed": "year and authors set to the paper; Cellosaurus provenance moved into the title"
}
the false defect
{
  "pmid": "34693799",
  "epubdate": "2021 Oct 23",
  "pubdate": "2023 Jan",
  "reading": "the entry's stored 2021 is correct; my checker compared pubdate alone"
}
placeholder blocks across both axes
{
  "pmid_side": 49,
  "doi_side": 55,
  "total": 104,
  "reading": "blocks where nobody transcribed an author list, so nobody compared the record to the claim; the subset where a wrong identifier would survive longest, and the only population where these checks have any power"
}
Still open
  1. Apply the two-date fix to the DOI check. Crossref's single `issued` field produces 1,294 off-by-one blocks that cycle 123 already read and found legitimate; the reading was correct and the instrument was never adjusted, so a standing checker will cry wolf on ~1,300 citations every run.
  2. Read the 104 placeholder blocks across both axes. They are the only population where these checks have power - everywhere else the stored value came from the registry being compared against - and they have now been identified twice and read zero times.
  3. Check whether the one defect found is representative. It was in a counter-evidence citation, the least-scrutinised part of an entry; if error rate tracks scrutiny, primary-claim citations and counter-evidence citations are different populations and neither check separates them.
  4. Ask whether the cited paper supports the claim. Six cycles of citation hygiene have established that the identifiers are right and have said nothing about whether the papers say what the entries claim, which is the only question that affects the biology.
Every source behind this entry (1)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.analysis

    Walking every committed node file for citation blocks carrying a pmid gives 17,925 blocks across 3,663 distinct PMIDs. Fetching each record from PubMed esummary in batches of 180 and comparing: 17,755 blocks state a comparable year, of which 17,748 match either the record's issue year or its online-first year exactly, 100.0 percent, with 7 off by one and 0 disagreeing by two years or more; 16,281 blocks state a comparable author list, of which 16,232 share at least one surname with the record, 99.7 percent, and 49 share none. All 49 store a placeholder rather than an author list. Before accepting the online-first date, two blocks disagreed by two years or more: PMID 34693799, epub 2021-10-23 and issue 2023 Jan, where the entry's stored 2021 is correct, and PMID 7759961 in line-u937-calm-af10-adult-monocytic, stored as year 2026 against a record of Reid YA, Journal of Leukocyte Biology, 1995.

    What this evidence cannot show

    AGREEMENT IS NEAR-TAUTOLOGICAL FOR MOST BLOCKS, as on the DOI side: these citations were largely written by fetching the record at the time, so the stored values came from the registry this compares them against, and the check has real power only over the blocks written from recall, which cannot be identified. THIS DOES NOT RE-VALIDATE THE INSTRUMENT - the two documented wrong-paper failures were corrected long ago and appear in the corpus as confessions, so their absence here is expected and means nothing; cycle 124 asked the validation question and got 16 of 23. SURNAME OVERLAP IS NOT IDENTITY and one shared surname passes a block. 1,644 blocks store no comparable author list and are outside that axis. THE 100.0% YEAR FIGURE WAS REACHED BY LOOSENING THE COMPARISON to accept either the issue or the online-first year, a change made AFTER seeing which cases failed - defensible because online-first is a real publication date, but it is the second number this instrument produced. The tokeniser is reused from the DOI side where it was debugged three times by reading worklists.

    Evidence that argues against this (3)
    • A 100.0% year rate should increase suspicion rather than confidence. It was obtained by accepting two different values as correct and by fixing the single block that disagreed, which between them guarantee the result. The defensible statement is that no PMID citation in this corpus disagrees with its record on a year that both PubMed fields admit - a much narrower claim than 'the years are right'.
    • Finding one defect in 17,925 blocks may reflect how these citations were created rather than how careful anyone was. They were written by agents instructed to fetch the record before writing, so the corpus is close to a copy of PubMed's metadata and this check largely verifies that a copy is a copy. It says almost nothing about whether the cited paper supports the claim it is attached to, which no check in this project touches.
    • The one defect found was in a COUNTER-EVIDENCE citation, not in a primary claim - the least-read part of an entry. If error rate correlates with how closely a block is scrutinised, then a single defect in the counter-evidence of one entry is weak evidence that the primary citations are clean, and the sampling here cannot distinguish the two populations.
Connected to