No preprint in this graph is carried at strong or established - and the audit that found that first misclassified Genes & Development as a preprint
also called preprint exposure · 10.1101 trap · peer review audit
Thirty-two citations here rest on work that has not been peer reviewed, none of them at the two highest confidence levels, and only one failed to say so. My first attempt reported 132, because the web-address prefix used by the main preprint server is also used by several respected journals from the same publisher.
The same thing, for a scientist
First audit of unreviewed sources. **32 evidence items rest on preprints; 0 are carried at `strong` or `established`** (20 moderate, 14 weak); 15 distinct preprints, mostly bioRxiv with 5 from Research Square. **One item did not disclose its preprint status in the caveat** and has been corrected. A second flagged item turned out to disclose correctly in different words - *'Conference abstract... CONFERENCE-SOURCED'* - which the regex did not match. **The finding is the trap.** `10.1101/` is Cold Spring Harbor Laboratory Press, which publishes **bioRxiv and medRxiv AND Genes & Development, Genome Research, CSH Perspectives, RNA and Learning & Memory.** Treating the prefix as a preprint marker classified peer-reviewed papers as unreviewed and inflated the count from 32 to **132**.
What the audit found
| evidence items resting on a preprint | 32 |
carried at established | 0 |
carried at strong | 0 |
| moderate | 20 |
| weak | 12 |
| distinct preprints | 15 |
| failing to disclose preprint status in the caveat | 1 |
The discipline holds. A confidence tier is a statement about how much weight a claim can bear, and an unreviewed single report cannot bear the top two. Nothing in this graph does. That is the result, and it was not guaranteed - three beats this month cited preprints, each under time pressure.
The one gap was concept-phase2-compute-backend-matching E2, citing a bioRxiv preprint at moderate with a caveat about benchmark conditions that never said the source was unreviewed. Corrected.
A near-miss worth recording: drug-bleximenib E6 was flagged as undisclosed and is not. Its caveat says 'Conference abstract... CONFERENCE-SOURCED (flagged 2026-09-19)'. The practice was right and the checker's vocabulary was wrong - the same shape as the counter-note detector penalising cross-referencing.
The trap: a DOI prefix shared by a preprint server and its publisher's journals
10.1101/ is Cold Spring Harbor Laboratory Press. It covers:
| DOI | what it is |
|---|---|
10.1101/2025.07.08.661805 | bioRxiv preprint |
10.1101/gad.280495.116 | Genes & Development, peer reviewed |
10.1101/gr.* | Genome Research, peer reviewed |
My first version matched ^10.1101/ and reported 132 preprint-sourced items, including Genes & Development and Genome Research papers that this graph relies on heavily. The real number is 32.
The discriminator is the shape, not the prefix: bioRxiv and medRxiv DOIs are date-formed - 10.1101/YYYY.MM.DD.NNNNNN - while the journals use a short journal code. That is now in data/symbol-collisions.json so an automated pass reads it.
This is the fourth instrument of mine this month to over-flag, after the gene queue reading ALL as a symbol, the contradiction screen reading scoping language as negation, and the accession audit inventing an accession from a fallback. The pattern is consistent enough to state as a working rule: an instrument built quickly against a heterogeneous corpus will find its own construction before it finds a defect.
- Do not read '0 preprints at strong or established' as a statement that the strong claims are sound. It says unreviewed sources are not over-weighted. The commoner failure in this graph is a peer-reviewed single paper generalised across lineage, which this audit does not touch.
- How many of the 15 preprints have since been published? Each should be re-cited to the journal version, and this audit cannot tell a stale citation from an unreviewed one.
- Is peer review the right line? The graph carries unreplicated peer-reviewed single papers at strong, which may be a weaker source than a well-powered preprint.
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis
Of the evidence items in this graph that cite a preprint, 32 exist and none is carried at strong or established confidence; 20 are moderate and 12 weak, drawn from 15 distinct preprints, and exactly one failed to disclose its preprint status in its caveat.
preprint-exposure: every citation in the graph classified by peer-review status, by venue string and DOI shape · Oncofusion knowledge graph (internal analysis) 2026What this evidence cannot show
Run 2026-09-20 against 1,555 entries. **This is a FLOOR, not a census**: detection reads the venue string and the DOI shape, so a preprint recorded with a blank venue and a non-preprint DOI is invisible. **It also cannot tell a genuinely unreviewed source from a stale citation** - a preprint since published in a journal but still cited by its preprint DOI counts here as unreviewed when the work has in fact been vetted. Moderate for both reasons; the zero at the top two tiers is the robust part, because those entries were checked by reading.
Evidence that argues against this (1)
- Zero preprints at the top tiers is a statement about where confidence labels were applied, not about whether the labels are right. This graph carries many single unreplicated peer-reviewed papers at strong - peer review is a weaker guarantee than a tier boundary implies, and drawing the line at peer review rather than at replication may be the wrong line to police. Shan X et al. 2017
COMPANION MEASURE: peer review is one reviewer-round; recency is about whether anyone has had time to challenge the result at all.
- refutes →Compute: which backend for which piece of phase-2 work, and the one constraint that matters for each
THE ONE DISCLOSURE GAP: it cited a bioRxiv preprint at moderate with a caveat that never said the source was unreviewed. Corrected 2026-09-20.
A NEAR-MISS THAT WAS RIGHT: flagged as undisclosed, it actually says 'Conference abstract... CONFERENCE-SOURCED'. The practice was right and the checker's vocabulary wrong.
THE OTHER HALF OF SOURCE INTEGRITY: nothing cited is retracted, and nothing unreviewed is over-weighted. Neither check touches the commoner failure - a single paper generalised.
FOURTH INSTRUMENT THIS MONTH TO OVER-FLAG: ALL as a gene symbol, scoping language as negation, an accession invented by a fallback, and now a shared DOI prefix.