A congress abstract book is one indexed document containing thousands of unrelated abstracts, so it matches almost any conjunction: across 18 Europe PMC full-text intersections read to completion, 83 of 321 records (25.9%) are proceedings, ranging 0% to 50% per query
also called Q115 proceedings contamination · abstract-book trap
Searching the whole text of papers is more powerful than searching summaries, but it has a new way of going wrong: a conference programme is filed as a single document containing thousands of separate short reports, so it matches nearly any combination of search words. A quarter of the results this beat got were conference programmes.
The same thing, for a scientist
Europe PMC `BODY:` search is the instrument this project now reaches for when rule 80 says an abstract-level bound is not a bound. It has a contaminant that abstract-level search does not. **A congress abstract book, poster session or meeting supplement is indexed as ONE document whose body contains hundreds to thousands of unrelated abstracts**, so any conjunction of terms that appear anywhere in a large haematology or oncology meeting matches it. Measured over the 25 probe intersections this beat ran: **741 records inspected, 94 proceedings (12.7%) overall**; restricted to the **18 queries whose entire result set was read** - the only honest denominator, because the other 7 were truncated at 60 of a larger set - **83 of 321 = 25.9%** [E1]. The rate is not uniform and the range is the useful number: **0% to 50%**. It is highest where a clause contains a word common in a meeting programme (`cerebellum`, `condensate`, `established disease`) and **zero where every clause is a specific compound or gene name** - `LEDGF x menin-inhibitor-list` returned 0 proceedings in 100 records read [E2]. Consequence: a `BODY:` hitCount is not a screening burden, it overstates it by up to half, and **every returned title must be read before a count is quoted.**
Worked cases
| intersection | hits | read | proceedings | rate |
|---|---|---|---|---|
| LEDGF x established-disease phrases | 12 | 12 | 6 | 50.0% |
| ENL alias group x cerebellum | 22 | 22 | 10 | 45.5% |
| MLLT1 x cerebellum | 16 | 16 | 7 | 43.8% |
| KDM2B x nucleolar x KMT2A | 27 | 27 | 10 | 37.0% |
| MLL fusion x phosphopeptide | 24 | 24 | 8 | 33.3% |
| ENL x condensate x phospho | 40 | 40 | 11 | 27.5% |
| menin-inhibitor list x gain vocabulary | 28 | 28 | 4 | 14.3% |
| WRAD x tiling CRISPR | 14 | 14 | 1 | 7.1% |
| H3K79 x decay rate | 7 | 7 | 0 | 0% |
| junction-antibody phrases x KMT2A | 3 | 3 | 0 | 0% |
| LEDGF x menin-inhibitor name list | 116 | 100 | 0 | 0% |
The recurring offenders are the same handful of documents: Abstract Book of the Congress of the European Hematology Association (several years), HemaSphere 'Publication Only' supplements, Haematologica Italian Society of Hematology congresses, SIOP scientific programmes, the Japanese Cancer Association annual meeting abstracts, ESHG poster collections and FEBS Open Bio 'Posters'. One of them, the EHA2024 congress book, appeared in six different intersections in this beat.
Why this is a different failure from the ones already on the books
Rule 94 ((A OR B)[tiab] silently drops the tag) and rule 95 (an unquoted translocation becomes a UID filter) both change which search runs. This one runs the search you wrote and returns documents that satisfy it. The match is real; the unit of indexing is wrong, because the document is a container rather than a work.
It is also the opposite sign from rule 80. Rule 80 says an abstract-level search is too narrow to bound a claim. Moving to full text fixes that and introduces a breadth error, so a beat that switches instruments trades a false negative for a false positive and should say which it is now exposed to.
And it is invisible at the top of the ranking. Europe PMC's relevance order puts proceedings last in every intersection here where the whole set was read; a beat that reads the first ten hits of a 40-hit query will see none of them and will still quote 40 as its burden.
What to do
- Read every title before quoting a
BODY:count, or state the count as an upper bound. - Report the proceedings fraction alongside the count. It is one regex over the titles and it is the difference between a burden of 40 and a burden of 29.
- Prefer clauses made of specific tokens. A compound name or a gene symbol in every clause drove the rate to zero here; an ordinary anatomical or cell-biological word drove it to 45%.
- Europe PMC has a
PUB_TYPEfacet, and rule 85 already records that its values must be quoted. Excluding proceedings by facet was not attempted by this beat and is the obvious next step.
| script | none - scratchpad q115_battery.py and a title regex; 2026-09-22 |
|---|---|
| status | run |
| full text corpus denominator | 12362145 |
| queries attempted | 25 |
| queries succeeded | 25 |
| queries failed | 0 |
| closes | 25 attempted = 25 succeeded + 0 failed |
| records inspected all queries | 741 |
| proceedings all queries | 94 |
| rate all PCT | 12.7 |
| fully read queries | 18 |
| records fully read | 321 |
| proceedings fully read | 83 |
| rate fully read PCT | 25.9 |
| detector | case-insensitive title match on abstract|congress|annual meeting|poster|proceedings|scientific programme|publication only|symposium|autorenregister |
| most promiscuous single document | HemaSphere, EHA2024 Hybrid Congress - appeared in six different intersections in this beat |
| not attempted | exclusion by Europe PMC PUB_TYPE facet; rule 85 requires the value to be quoted and no proceedings value was tested |
| relationship to existing entries | New. Rules 94 and 95 concern a query that does not execute as written; this concerns a query that executes correctly and returns a document type whose unit of indexing defeats the inference. It applies to every Europe PMC BODY: bound this project writes from now on, including the nine written by this beat. |
| absence class removed 2026 09 23 | This entry carried absence_class = 'not-an-absence -- this is an instrument characterisation'. It is not an absence record - it characterises an instrument, namely that a congress abstract book is one indexed document and therefore matches any conjunction. The field said so in words, and the right fix is to remove it rather than to invent a token. Recorded because it answers an open question raised in cycle 68: some of the 243 entries carrying absence_class are not absences at all, and the denominator of every absence statistic this project quotes should exclude them. |
per query range PCT
[ 0, 50 ]
repeat offenders
[ "HemaSphere EHA congress abstract books (2020, 2022, 2024) and 'Publication Only' supplements", "Haematologica, Italian Society of Hematology and EHA congresses (2011, 2013, 2014, 2016, 2019, 2020)", "Pediatr Blood Cancer SIOP scientific programmes (2015, 2017)", "Cancer Sci Japanese Cancer Association annual meeting abstracts (2018, 2021, 2023, 2026)", "Eur J Hum Genet ESHG poster collections", "FEBS Open Bio 'Posters'", "Indian J Hematol Blood Transfus ISHBT annual conferences" ]
- Does PUB_TYPE exclusion work? A quoted facet value that removes congress supplements would make every BODY: bound in this project cleaner at zero cost, and rule 85 already gives the quoting discipline.
- How many of this corpus's existing Europe PMC full-text counts were quoted without a title screen? Any count from a BODY: intersection written before today is an upper bound and nobody marked it as one.
- Is a congress abstract inside one of these books ever the ONLY published record of a result this graph wants? If so the books are not purely contaminant and should be screened rather than excluded - which is a much larger reading job.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database
Across 18 Europe PMC BODY: intersections whose complete result sets were retrieved and title-screened, 83 of 321 records are congress abstract books, poster sessions or meeting supplements - 25.9% - with a per-query range of 0% to 50%.
Europe PMC RESTful article search, 25 BODY: intersections run and title-screened 2026-09-22 (scratchpad q115_proceedings.json) · Europe PMC 2026What this evidence cannot show
Proceedings were identified by a case-insensitive title regex on abstract|congress|annual meeting|poster|proceedings|scientific programme|publication only|symposium|autorenregister. That will miss a proceedings volume with an idiosyncratic title and will falsely flag a genuine research paper with 'abstract' in its title, so the figure is approximate in both directions; every flagged title in the fully-read sets was eyeballed and none was a false positive. 25 queries attempted, 25 succeeded, 0 failed.
Evidence that argues against this (1)
- The pooled 25.9% is driven by the small queries, which are exactly the ones a beat can read completely, so it is a rate over the queries that matter rather than over records in general. Pooling all 741 inspected records including the truncated sets gives 12.7%, and that lower figure is the one that would be quoted by someone who did not separate the truncated queries. Europe PMC 2026
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database
The contamination is concentrated in queries containing a common word: 100 records read across the first two pages of the LEDGF x menin-inhibitor-name-list intersection (116 hits) contain zero proceedings, while the LEDGF x established-disease-phrase intersection is 6 of 12.
Europe PMC RESTful search, paged retrieval of BODY:(LEDGF aliases) AND BODY:(20 menin-inhibitor names), pages 1-3, 2026-09-22 · Europe PMC 2026What this evidence cannot show
Page 3 returned 0 records against a stated hitCount of 116, so 100 of 116 were read and 16 were not retrievable through paging; the zero therefore covers 86% of the set. This is itself worth noting as an endpoint quirk.
Evidence that argues against this (1)
- An alternative explanation is relevance ordering rather than composition: if proceedings always rank last, the 16 unretrieved records could be exactly the proceedings. That cannot be excluded from these data, and it would mean the rule is 'proceedings rank last' rather than 'specific tokens exclude them'. Both readings imply the same practice - read the tail, not just the head. Europe PMC 2026
characterises the instrument that census recommends for re-bounding the 23 abstract-level absences, and names the breadth error it introduces
the contamination measurement that qualifies every full-text count written by this beat