Research analytics: the data

Six files, shown as the model gets them in the conditions that read CSV. generate_data.py writes them from a fixed seed.

There are 52 designed works and 33 background works. The background works are noise: their authors are at four background institutions, and they cite only each other, except for the few that set up the citation-window questions.

The hierarchy: Tessel Laboratory (I04, LAB) sits under the Albury Centre for Marine Robotics (I03, LAB), under the University of Albury (I02, UNI), under the Albury University System (I01, SYS), which also holds Albury Metropolitan University (I05). Halloran Institute (I06, AFF) has I02 as its parent but is not a unit of it. Kesting Institute of Hydrodynamics (I08, LAB) sits under Kesting University (I07).

The traps these records set, and the questions that test each, are on the Questions page.

The R10 executed condition gets the tables execute.py writes: - test records dropped; - preprints merged into their published versions; - each work flagged as research output or not; - authors merged into people; - one credited institution per person and work; - the institution roll-ups; - tiers ranked.

The base data plus 765 invented works, from a fixed seed, so that the same 48 questions need counting and joining across many records. domains/research/scale.py builds it, and the keys are computed from the data, as for the base.

The base records stay as they are, so every question keeps its point. 515 of the added works repeat the traps, many of them on the designed people, institutions and sources, and the other 250 are background:

• Who is who: Mara Ellison gets a fourth record, and a third, different "M. Ellison" works at Rostval. Author records that share an ORCID go from 3 to 12, and names on more than one record from 2 to 9.
• Hierarchy: new labs under the University of Albury, Albury Metropolitan and Rostval, and two new affiliated institutes (AFF), under Albury Metropolitan and Kesting.
• Affiliations: authorships with several affiliations go from 2 to 26, in both orders, so the rule about the last one decides many credits.
• Versions: preprints with a published version go from 2 to 23, some under another record of the same person.
• Test records, retractions and delistings: STG works go from 3 to 25, retracted works from 1 to 13, and works in a K0 source from 2 to 24.
• Document types: more supplements (SUP), editorials, errata and front matter.
• Citations: about 1,000, with new citing works at every level of self-citation, and blocks that keep each comparison's answer: Kesting and Rostval still swap places between full and fractional counting, and W30 still leads in total while W31 leads within its window.

Of the 46 keys graded by code, 31 change from the base data's. The 15 that don't are comparisons whose answers hold by design, or questions about one record.

The R2 rdf and R6 rdf+ontology prompts are about 770 KB, against 136 KB for R1 csv, which is too large for the model's context, so the runs on this dataset leave them out.