Replication archive for "When the Rubric Answers Its Own Hard Case: Auditing Rule-Installed Conventions in Text-Derived Indices"

Data, coding instruments and analysis code for a study of what a text-derivedpolicy index records when its rubric settles its own hard case. The archive reproduces every quantity reported in the manuscript. Its layoutmirrors the working directory, so the scripts run in place: analysis scripts sitat the root and read data/ and results/ beside them. Contents- pilot/ : the coding rubric in each of its 7 versions, including the neutral-round rubric with the adjudicating clause deleted- instruments/ : the coder-facing forms for the blind rounds and the panel- data/ : record-level codings for every round, human and machine (19 files)- results/ : the 29 canonical result files the manuscript reads- *.py : the 85 analysis scripts that produce those results- panel_returns/ : four independent panel readers' returns, under coder labels- STOPPING_RECORD.md : when and why panel recruitment stopped, written before the result was known De-identificationThe panel's returns carry coder labels, not names, and the label-to-name mapping isheld privately and is not deposited: the panellists consented to publication of theircodings in de-identified form only. The two blind rounds were coded by two differentpeople, which the manuscript states; their coding files are named by round rather thanby person. Because the mapping is withheld, the manuscript cannot be re-knitted fromthis archive alone, though every number it reports is reproducible from results/ andevery result file is reproducible from the scripts and data/. Source dataThe corpus is the Global Trade Alert database, publicly available from its maintainers.Codings are keyed on GTA intervention id, so the source text can be re-fetched. Theblind-round and negative-control files carry the truncated description text becausethose texts are the instrument, and a coder round cannot be replicated without theexact stimulus. GTA material remains subject to its own terms of use. Data, coding instruments and analysis code for a study of what a text-derivedpolicy index records when its rubric settles its own hard case. The archive reproduces every quantity reported in the manuscript. Its layoutmirrors the working directory, so the scripts run in place: analysis scripts sitat the root and read data/ and results/ beside them. Contents- pilot/ : the coding rubric in each of its 7 versions, including the neutral-round rubric with the adjudicating clause deleted- instruments/ : the coder-facing forms for the blind rounds and the panel- data/ : record-level codings for every round, human and machine (19 files)- results/ : the 29 canonical result files the manuscript reads- *.py : the 85 analysis scripts that produce those results- panel_returns/ : four independent panel readers' returns, under coder labels- STOPPING_RECORD.md : when and why panel recruitment stopped, written before the result was known De-identificationThe panel's returns carry coder labels, not names, and the label-to-name mapping isheld privately and is not deposited: the panellists consented to publication of theircodings in de-identified form only. The two blind rounds were coded by two differentpeople, which the manuscript states; their coding files are named by round rather thanby person. Because the mapping is withheld, the manuscript cannot be re-knitted fromthis archive alone, though every number it reports is reproducible from results/ andevery result file is reproducible from the scripts and data/. Source dataThe corpus is the Global Trade Alert database, publicly available from its maintainers.Codings are keyed on GTA intervention id, so the source text can be re-fetched. Theblind-round and negative-control files carry the truncated description text becausethose texts are the instrument, and a coder round cannot be replicated without theexact stimulus. GTA material remains subject to its own terms of use.