MétaCan

1,000 works · Opus 4.8 · GPT-5.6 (high) · Grok 4.5 · one locked rubric

The field’s boundary is not a line. It is a region.

Three frontier models screened the same 1,000 works, drawn from the real frame with known selection probabilities, against the same locked rubric on its full eight-field payload. Of the 51 works any model called metaresearch, only 19 (37.3%) were called metaresearch by all three, and 24 (47.1%) rest on a single model's opinion.

Two screeners can agree on a rate while finding almost entirely different works. At a ~1% base rate, the settled rejects buy 98% agreement for free — which is why an agreement statistic computed over the whole sample tells you nothing about the boundary, and why the table below, not a percentage, is the deliverable.

Any model said metaresearch
51
the disagreement dossier: the field's empirical boundary
All three agreed
19
37.3% of the dossier — the settled core
Two of three
8
contested
One model only
24
47.1% of the dossier rests on one model's opinion

The three models are not interchangeable

How many of the 1,000 works each model called metaresearch (tier T1 or T2), on identical input. T3 is adjacent and does not count as in scope, which is why a model's count can never exceed the 51 works in the dossier.

Claude Opus 4.8
42
GPT-5.6 (high)
31
Grok 4.5
24

Swap which model you call “the screener” and the size of the field moves. That spread — not the binomial confidence interval on any one model's labels — is the honest uncertainty on how big Canadian metaresearch is.

The disagreement dossier

All three verdicts, confidences and reasons side by side. These are the works against which the inclusion criteria have to be written — because these are the works on which reasonable screeners, given the same rubric and the same evidence, disagree.

8 works
WorkStratumn_inOpus 4.8GPT-5.6Grok 4.5
La méthodologie au service de l'avancement des connaissances en matière de jeux de hasard et d'argent
2012 · article · fr · Journal of Gambling Issues
french
w=1554
2/3T1low
conceptual

Reviews longitudinal, experimental, and genetic study designs and what they enable for advancing gambling knowledge; the object is research methodology and study design, though it is field-oriented.

T1high
conceptual

This explicitly examines research methodologies used to advance knowledge in gambling studies.

OUTmedium
conceptual

Reviews longitudinal, experimental, and genetic methods as tools to advance gambling knowledge; domain methods guidance, not study of research itself.

Représentations mentales maternelles : concepts et mesures liés à l’attachement
2009 · article · fr · Enfance
french
w=1554
2/3T2low
conceptual

Critical review of the concepts and measures used in the attachment literature, explicitly flagging insufficiently demonstrated validity and weak empirical rigor; sits on the boundary between domain theory and a critique of measurement practice.

T1medium
conceptual

The article evaluates concepts and measurement approaches used in attachment research, making research measurement its primary object.

OUTmedium
conceptual

Domain psychology review of attachment-related constructs and measures; advances developmental theory rather than studying research as practice.

Using Logic Models and the Action Model/Change Model Schema in Planning the Learning Community Program: A Comparative Case Study
2018 · article · en · Canadian Journal of Program Evaluation
venue_new
w=2684
2/3T1low
empirical

Comparative case study assessing the relative strengths of logic models versus the action model/change model schema as planning and evaluation tools; the object is evaluation methodology, which sits on the boundary between applied evaluation practice and research methods.

T1medium
empirical

This comparative case study evaluates program-evaluation models and their properties as methodological tools.

OUTmedium
empirical

Compares logic models for planning a learning-community program; program evaluation methods, not evaluation of research.

Canadian Supercomputer Threat Assessment and Potential Responses
2019 · article · en · The Journal of Intelligence Conflict and Warfare
venue_new
w=2684
2/3T2low
policy

Briefing note assessing espionage and security threats to Compute Canada's national supercomputing infrastructure; object is the governance of Canadian research infrastructure, though it sits on the T2/T3 boundary as a policy briefing.

T2high
policy

The briefing analyzes security risks and responses involving Canadian research supercomputing infrastructure.

T3medium
policy

Policy briefing on security threats to Canadian research supercomputers, relevant to the research ecosystem but not analytic metaresearch.

The emergence of politics as a taught discipline at universities in the United Kingdom
2019 · article · en · The British Journal of Politics and International Relations
fund_new
w=1679
2/3T2medium
empirical

Archival history of how political science became a taught discipline in UK universities, tracing chairs, subject associations and curricula; the object is an academic discipline's formation.

OUThigh
empirical

It reconstructs the history of political-science teaching and curricula, not research practice.

T2medium
empirical

Historical study of politics as a taught university discipline; adjacent history of the academic field.

Most Recent Data (2005). Statistics Canada. CANSIM: Science and Technology - Research and Development | Country: Canada | Table: Survey of innovation, logging and manufacturing industries, percentage of plants with expenditures on research and development services by different locations | Variable: From all other countries, Clothing manufacturing and leather and allied product manufacturing, 0% of expenditures, Innovative plants | Units: %, 2005. Data-Planet™ Statistical Ready Reference by Conquest Systems, Inc. Dataset-ID: 075-001-182.
2015 · other · en · Data Planet
about_only
w=3321
2/3T1medium
empirical

Statistics Canada CANSIM table of research and development expenditure by Canadian innovative plants; the object is R&D activity and its funding in Canada, though the record is a repackaged statistical table rather than a study.

T2high
empirical

The dataset documents Canadian research and development activity as part of the national research system.

OUThigh
other

Packaged industrial innovation R&D expenditure statistic for manufacturing plants, not a study of research practice.

Structural equation modeling in the plant sciences: An example using yield components in oat
2011 · article · en · BioOne Complete (BioOne)
about_only
w=3321
2/3T1medium
conceptual

Tutorial introducing plant scientists to structural equation modeling, reviewing its principles and offering guidelines on when it is appropriate; the oat field trial is an illustrative vehicle and the object is statistical practice in a research community.

OUThigh
empirical

Structural equation modeling is used to analyze oat yield, so the method is not the object of study.

T1medium
empirical

Primary aim is introducing SEM as statistical practice for plant scientists, with guidelines on when SEM is appropriate.

Factores asociados al uso de tácticas políticas en el ascenso laboral: una evaluación desde el contexto de la Universidad de Los Andes, “Núcleo Dr. Pedro Rincón Gutierrez”, Venezuela.
2019 · article · es · DOAJ (DOAJ: Directory of Open Access Journals)
about_only
w=3321
2/3T2low
empirical

Survey of factors behind the use of political tactics in academic promotion among university professors in Venezuela; grounded work on the academic workforce and careers, but framed as organizational behavior.

T1medium
empirical

It studies political tactics in university professors' career advancement, making the academic research workforce the object.

OUTmedium
empirical

Organizational-politics survey of professors’ promotion tactics; object is workplace behaviour, not research practice or the research workforce as such.

What the screen actually found

Three frontier models (Opus 4.8, GPT-5.6 high, Grok 4.5) screened the same 1,000 works from the REAL 4.3M frame, on the rubric's FULL eight-field payload, with randomized manifest-logged chunks and harness-written labels: every defect D1, D2, D11 and finding 16 identified, repaired. Design-weighted base rates span 1.89% to 3.54% (1.9x). But the sets are the finding, as finding 16 predicted: of the 51 works ANY model called metaresearch, only 19 (37%) were called metaresearch by ALL THREE, and 24 (47%) rest on a single model's opinion. THE FIELD'S BOUNDARY IS NOT A LINE THE MODELS SHARE; IT IS A REGION THEY EACH CUT DIFFERENTLY. GPT-5.6 also violated the locked output schema on 18 of 1,000 records, writing genre values into the tier field, which the manifest validator caught. The deliverable is not the base rate: it is the disagreement dossier, the 51 works that mark the empirical boundary of the field and against which the inclusion criteria must actually be written.

Computed 2026-07-12T19:52:14Z · all 22 findings · this dossier as JSON