MétaCan

Landscape

What the machine-labelled subset of the frame looks like: categories, study designs, years, languages. Below it, the frame described by itself.

Read the coverage before the counts. Labels cover 700 of the 4,299,418 works in the frame (0.016%). Every count in this section is over that labelled subset only; it says what the labelled works look like, never how much of the frame is in a category. These are machine labels (frontier LLM, unvalidated), and an unlabelled work is NOT a negative.
Labelled works
700
works with at least one model label
Label rows
2,000
one per (work, model) pair
Models
gpt · grok · opus
each work is labelled by up to three

Labelled works by category

A work counts under a category if at least one model applied it; the darker figure requires every model that labelled the work to agree. The gap between the two columns is the models disagreeing, and that gap is a finding, not noise.

CategoryAny modelAll models agree
Metaresearch11635
Insufficient payload (model declined to judge)6823
Meta-epidemiology (broad)515
Scholarly communication236
Science and technology studies184
Open science106
Research integrity52
Bibliometrics32
Meta-epidemiology (narrow)21

Labelled works by study design

Same two readings: any model, and all models in agreement. No design label here is MEDLINE-validated yet; when that validation lands it will be marked explicitly.

Study designAny modelAll models agree
Not applicable203111
Other design19643
Theoretical or conceptual13758
Observational12365
Qualitative9757
Systematic review7443
Simulation or modeling7239
Bench or experimental4227
Meta-analysis156
Randomized trial75
Non-randomized trial31
Case report11

Labelled works by year

Where the labelling rounds have reached so far. This is coverage of the label table, not a property of the field.

YearLabelled works
20006
20015
200210
200311
20047
200515
20069
200715
200820
200914
201011
201131
201225
201326
201422
201534
201622
201752
201823
201926
202046
202149
202242
202365
202469
202545

Labelled works by language

Coverage again: the labelling rounds sample the frame, and the frame is 6% French.

LanguageLabelled works
en645
fr34
unknown12
es3
lv1
de1
id1
cs1
it1
nl1

The frame itself

Everything below is over all 4,299,418 works, labelled or not. Every figure is a query against the database, computed at request time and cached for an hour; nothing here is a number someone typed.

Works in the frame
4,299,418
No Canadian affiliation
1,565,226
36.4% of the frame
No abstract
1,003,117
23.3% of the frame
Retraction notices
1,052
joined from Retraction Watch

Works by year

The frame over time, with the works that carry NO Canadian affiliation drawn underneath. The gap between the two lines is what an affiliation-only frame silently loses — 36.4% of the frame, 1,565,226 works.

Works by route

Why each work is in the frame. The four routes OVERLAP — a work can be admitted by several — so these bars sum to 5,476,864, which is 1,177,446 more than the 4,299,418 works in the frame. That overlap is the next chart.

The overlap: exact route combinations

Each work counted once, under the exact set of routes that admitted it. Teal bars are works admitted by a SINGLE route: remove that route from the design and those works vanish from the frame entirely.

Works by field

OpenAlex's primary field, as recorded.

Works by language

French is highlighted. It is 6% of the frame, it is oversampled in the screen on purpose, and it is the language the abstract cascade rescues worst (15.4% recovery against 38.8% for English).

The abstract gap is structural, not noise

Share of works with NO abstract, by type, worst first. 23.3% of the frame has no abstract, and the screen finds HALF as much metaresearch there. If the gap were random, a better index would fix it. It is not random: it is concentrated in types that never carry an abstract at all — so "just screen the works that have abstracts" is a selection on a covariate that predicts the outcome.

The post-publication record has four states, and OpenAlex has a boolean

1,052 works in the frame carry a Retraction Watch notice. The solid bar is what OpenAlex flags; the hatched bar is what it reports as “false” — 143 works whose notice OpenAlex does not carry, which a reader takes to mean “fine”. An expression of concern is not a retraction, and `is_retracted` has no way to say so.

StateWorksOpenAlex flags itOpenAlex reports false
Retraction95590649
Expression of concern52052
Correction34232
Reinstatement11110

Top venues

By work count in the frame.

Top funders

Split from the semicolon-separated funder string. The funder route admits works that carry no Canadian affiliation at all.

Teacher spread over the frame (provisional baseline)

Every one of the 4,299,418 works carries two provisional teacher-head scores, and the spread is how far the two heads sit apart on one work. These figures come from pilot/results/frame_scores.json, the file the scoring run writes; nothing here was typed by hand.

Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.

Works scored
4,299,418
every work in the frame, by the two-teacher panel
Mean teacher spread
0.2432
the average disagreement between the two heads
Teacher spread, 99th percentile
0.3963
99% of works sit below this spread
Works where the teachers would split
540
spread above 0.5

Every series here is available as JSON: see the API.