A quarter of this graph's load-bearing evidence is under twelve months old, and the year field is right in a way a checker would call wrong
also called citation recency · year drift · online-first versus issue date
About a quarter of the evidence carrying the most weight here was published within the last year, and no check this project runs would flag that - a paper from last month passes every one. It is partly unavoidable, because the drugs this graph is about reached patients while it was being written.
The same thing, for a scientist
**Recency is a fragility none of this project's checks can see.** Of 6,751 dated external citations, **24% are from 2026 and 29% from 2025 onward**. Restricted to items graded `strong` or `established`, **22% cite work under twelve months old**, and **146 of 1,477 entries have their entire top-tier external evidence from the last two years**. Separately, a graph-wide check of the recorded `year` against Europe PMC: **33 of 2,157 resolved PMIDs disagree (1.5%), and 97% of those are off by exactly one.** On inspection that is not error - it is the **electronic-versus-print publication date** distinction. **`verify-citations.py` does not check year, and this analysis concludes that omission is correct**: adding the check would produce 33 false positives.
What the existing checks cannot see
This project's citation discipline is unusually strict, and each check answers a different question about a source:
| check | question |
|---|---|
verify-citations.py | does it exist and match? |
check-retracted-sources.py | has it been withdrawn? |
preprint-exposure | was it peer reviewed? |
All three pass for a paper published last month that no one has attempted to replicate. Peer review is one reviewer-round; replication is the thing that actually hardens a result, and it takes years the calendar has not yet supplied.
The exposure, external literature only:
| count | share | |
|---|---|---|
| dated external citations | 6,751 | |
| from 2026 | 1,602 | 24% |
| from 2025 onward | 1,927 | 29% |
| load-bearing (strong/established) | 5,150 | |
| of those, from 2026 | 1,130 | 22% |
| entries whose top-tier evidence is entirely from the last 2 years | 146 | 10% |
The graph's own internal analyses were excluded - 291 items dated 2026 because this project wrote them this year. Including them would have inflated the 2026 share to 27% and made the finding partly an artefact of the graph citing itself.
This is exposure, not error. Menin inhibitors reached the clinic during this project's lifetime; for large parts of the therapeutic argument a 2026 trial read-out is the only evidence and is better than a 2005 cell-line paper. The point of measuring it is that the fragility should be stated rather than discovered - and that an entry resting entirely on twelve-month-old work is one replication failure away from needing a rewrite.
The year field is right, in a way a naive checker would call wrong
A 25-item sample turned up two year mismatches, which at 8% would have been alarming. Resolving every PMID that carries a recorded year: 33 of 2,157 disagree, 1.5% - and 32 of the 33 are off by exactly one year.
That distribution is diagnostic, and reading the outlier settles it. PMID 34693799 is recorded here as 2021 while Europe PMC's pubYear says 2023. Both are right:
firstPublicationDate: 2021-10-23 (electronic)
pubYear / printPubDate: 2023 (issue)
A journal published it online in October 2021 and put it in a 2023 print issue. The graph recorded the electronic date; the index reports the issue year. The two-year gap made it look like the worst case and it is in fact the clearest illustration.
verify-citations.py does not check year. That looked like a gap, and it is not: a year check would have produced 33 false positives and trained agents to "correct" values that are already right. The omission is the correct design, and it is now recorded as a decision rather than left as an accident.
What follows for practice: a year in this graph should be read as a publication year, not the one. Where a year is load-bearing - priority claims, 'the first report of', anything about what was known when - cite the electronic date explicitly and say which you used.
- A quarter of the load-bearing evidence being under a year old matters most where this graph touches treatment. A 2026 trial read-out is the best available evidence and also the least tested; anyone using this graph to argue for a clinical decision should check whether the supporting entry rests on work old enough to have been challenged.
- Do not 'fix' the 33 year mismatches. Most are electronic publication dates and are correct; changing them to issue years would destroy information, and a year check added to verify-citations.py would generate 33 false positives.
- Should entries whose top-tier evidence is entirely under two years old carry a standing caveat saying so? 146 entries is too many to annotate by hand and the flag is mechanical.
- Which of the 33 year mismatches are electronic dates and which are errors? Resolving each one's firstPublicationDate would settle it and was not done.
- Is there a check for the thing none of these measure - a result that has not been replicated, as distinct from one that is recent, unreviewed or withdrawn?
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis
Of 6,751 dated citations to external literature in this graph, 24% were published in 2026 and 29% in 2025 or later; restricted to evidence items graded strong or established, 22% cite work published within the last twelve months, and 146 of 1,477 entries carrying top-tier evidence have all of it from the last two years.
citation-recency-fragility: publication-year distribution of cited work, split by confidence tier and by internal versus external source · Oncofusion knowledge graph (internal analysis) 2026What this evidence cannot show
Run 2026-09-20. **The year is read from what each citation records**, which the companion check shows is 98.5% consistent with the index - so the histogram is reliable at the percent level and not at the item level. **Recency is exposure, not error**, and in this field it is frequently unavoidable. **The 146 figure counts only items with a usable year graded strong or established**; an entry also carrying older moderate evidence still appears, so it overstates how isolated those entries are. Moderate for that reason.
Evidence that argues against this (1)
- A single publication year cannot distinguish a result reported once from one reported and independently replicated within the same year, and this field has several fast replications - the BENC enhancer result, for instance, has independent genetic support from a different laboratory. Counting years treats those identically, so the measure overstates fragility wherever replication has been quick. Bahr C et al. 2018
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.analysis
Of 2,157 distinct PMIDs in this graph carrying a recorded publication year, 33 (1.5%) disagree with Europe PMC's pubYear, and 32 of those 33 differ by exactly one year, a distribution consistent with the distinction between electronic and print publication dates rather than with transcription error.
year-drift check: every recorded citation year resolved against Europe PMC in batches, with a hand-verified control PMID · Oncofusion knowledge graph (internal analysis) 2026What this evidence cannot show
Run 2026-09-20 with a control PMID whose year was verified by hand earlier in the session, per brief rule 67. **The interpretation rests on reading ONE outlier** - PMID 34693799, electronic 2021 and issue 2023 - and on the off-by-one distribution; the other 32 were not individually resolved to their electronic dates, so 'consistent with' is the strongest claim available. A transcription error would also usually be off by one.
Evidence that argues against this (1)
- Off-by-one is also exactly what a careless transcription looks like, so the distribution alone does not exclude error - it is the outlier's electronic and print dates that make the online-first reading concrete, and one confirmed case cannot establish the mechanism for the other thirty-two. Shi J et al. 2013
THE THIRD QUESTION ABOUT A SOURCE: exists and matches, not withdrawn, peer reviewed - and all three pass for a paper nobody has tried to replicate.
COMPANION MEASURE: peer review is one reviewer-round; recency is about whether anyone has had time to challenge the result at all.
WHY verify-citations.py OMITS THE YEAR, and why that is right: a year check would produce 33 false positives, because the recorded value is often the electronic date and the index reports the issue year.