Data and code for: Exact Loss Envelopes for Risk-Controlled Updating of Institutional Outcome Estimates: Evidence from Public Graduate Outcomes

This data and code archive accompanies the manuscript "Exact Loss Envelopes for Risk-Controlled Updating of Institutional Outcome Estimates: Evidence from Public Graduate Outcomes". It contains attributed institutional and programme-level aggregate numerical inputs from US PSEO, UK LEO, Korean KESS/KEDI and Chinese institutional graduate-outcome reports; frozen candidate predictions and interfaces; controller parameters and formulas; S/T result tables; China auxiliary calculations; envelope and norm-comparator analyses; budget diagnostics; and Python/R reproduction programs with their plotting source tables. The portable workflow reconstructs the update-stage calculations from frozen predictions without paid model APIs, candidate retraining or model weights. README.txt gives the software dependencies and execution commands. Core and archive-wide manifests record file sizes and SHA-256 checksums. Source-specific provenance, provider acknowledgements and licence records accompany the numerical data. No individual student records, complete source publications, manuscript files, author photographs or editorial correspondence are included. Licensing and reuse The MIT License applies to original software authored by Yuchen Wang, Lei Wang and Mengdi Yao, including the reproduction code and modular assembly helper, to the extent that the authors hold the relevant rights. Original documentation and original research results controlled by the authors are licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). These licences apply to their respective components; they are not alternative licences for the entire collection. Third-party data, source-derived material and third-party software retain their original licences, attribution requirements and other applicable terms. Neither MIT nor CC BY 4.0 replaces those terms or grants rights that the authors do not hold. Consult DATA_USE_AND_REUSE_STATEMENT.txt, SOURCE_ATTRIBUTION_AND_NOTICES.txt, LICENSES_AND_SCOPE.txt and the component-level notices for the applicable source conditions. MIT License: https://opensource.org/license/mit. CC BY 4.0: https://creativecommons.org/licenses/by/4.0/. Package contents The eight ZIP files are organised by scientific role. Together they contain all 781 files of the frozen Version 1.0 collection, with each original file path included exactly once. 01_Documentation_Code_and_Provenance.zipOriginal guides and statements; reproduction programs and plotting inputs; parameter records, data dictionaries, provenance, licences and integrity manifests. 02_Cleaned_Public_Data.zipCleaned aggregate numerical data from the public sources represented in the frozen collection. 03_Frozen_Model_Interfaces.zipFrozen model-interface files used by the reproduction workflow. 04_Frozen_Candidate_Predictions.zipFrozen candidate predictions for replaying the update-stage calculations. 05_Result_Tables_and_Risk_Summaries.zipMain and supplementary numerical source tables, final results, supplementary analyses and risk summaries. 06_Envelope_Ablation_and_Runtime.zipDirectional-envelope ablation results, runtime evidence and their numerical source tables. 07_Norm_Comparators_and_Mechanisms.zipNorm-comparator and mechanism results, associated views and validation records. 08_Budget_Regime_Diagnostics.zipBudget-regime diagnostics and associated numerical results. Download and verification For complete reproduction, download all eight ZIP files. Each is an ordinary ZIP that can be opened separately. Extract them into the same parent directory and merge their common Exact_Loss_Envelopes_Data_and_Code_v1.0 folder. Preserve all internal paths. The original data, programs and file contents are unchanged. START_HERE_Modular_Download.txt explains downloading, extraction and verification. ASSEMBLE_AND_VERIFY.py is an optional Python tool that assembles the eight packages and checks all 781 files. It uses only the standard library and does not access the network or calculate research results. PACKAGE_CONTENTS.csv records each file's package, path, size and SHA-256 checksum. SHA256SUMS.txt provides SHA-256 checksums for the eight ZIP files and the three supporting files above. The separately displayed README.txt and DATA_USE_AND_REUSE_STATEMENT.txt are reading copies of the unchanged files inside package 01. Consult them for reproduction commands, software requirements, source attribution and reuse conditions. Methods and interpretation The reproduction workflow starts from the retained candidate predictions, model interfaces and calibration/evaluation role records. It reconstructs the update-stage calculations and comparisons; candidate retraining, model-weight distribution and complete re-extraction of native source publications are outside its scope. The supplied programs cover model-interface reconstruction, same-period finite-pool (S) and temporal (T) controller replay, the auxiliary China calculation, directional-envelope and norm-comparator analyses, and budget, cost and runtime diagnostics. S reassigns roles within previously inspected finite pools. Repetitions do not add independent observations. T evaluates retained later outcomes empirically using the recorded calibration roles. The China summaries weight institutions equally after within-institution aggregation. Users should retain the supplied analysis units, endpoint definitions, denominators and observation windows when interpreting the results. Institution identifiers should be read as strings. Missingness, suppression and structural-inapplicability codes must be preserved and must not be converted to observed zeros. US industry composition, UK/Korean employment measures and Chinese report-defined graduate destinations are distinct endpoints. They should not be pooled as one harmonised employment rate or an unqualified cross-country loss average. The field dictionaries, parameter/formula registers and source records document the task-specific definitions. Reproduction environment and workflow The recorded Python replay environment is Python 3.9.6 with NumPy 2.0.2, pandas 2.3.3 and SciPy 1.13.1. These dependency versions are retained in requirements.txt and the accompanying environment records. Controller replay uses NumPy and pandas; the auxiliary China calculation also uses SciPy. Historical training-library records describe the original method context and are not all required for frozen-prediction replay. Figure reproduction uses Rscript and the dependencies listed in R_PACKAGES.txt: ggplot2, dplyr, tidyr, readr, patchwork, grid (base R), scales, svglite, ragg, jsonlite, digest, png, xml2 and ggrepel. The editable Figure 1 SVG also uses Inkscape. The retained figure guides record Arial font use and the corresponding plotting environments. Numerical values are preserved; layout, margins and guide positioning may differ between retained renderers. After assembling the eight packages, run the commands in the original README.txt from the common collection root. Keep the internal directory structure and send all rebuilt outputs to a new directory outside the collection. The usual sequence is integrity verification, interface reconstruction, controller replay and summary, followed by the required auxiliary analyses, numerical tables and figures. The controls and summary commands must use the same output directory. The --tracks S and --tracks T options select individual replay tracks. Integrity checks serve different scopes. ASSEMBLE_AND_VERIFY.py checks all 781 distributed original file paths and their content hashes. The original Run_Reproduction_运行复现.py --task verify --scope core checks the 304 frozen-core objects. Its --scope full uses the archive-specific inventory, which covers every archived file except that inventory itself. SHA256SUMS.txt checks the modular ZIP containers and their supporting delivery files. Container checks and extracted-file checks should both pass before reproduction. The numerical materials are supplied mainly as CSV and gzip-compressed CSV, with JSON metadata and TXT/Markdown documentation. Python and R source programs, a data-source-register PDF, an editable SVG source and a numerical TeX table fragment retain their documented roles. No particular graphical interface or operating-system-specific absolute path is required by the modular assembly helper; it uses only the Python 3 standard library and operates without network access. Provenance, attribution and citation The eight-package distribution preserves the frozen Version 1.0 collection. Original parameter identifiers and historical source/table locators remain unchanged. The author-archive:/ prefix identifies an original source-relative object; it is not a local path that users must recreate or an additional downloadable file. SOURCE_FILE_MANIFEST.csv records the source-to-archive mapping and the corresponding checksums. Provider release identities, access records, transformations and source-file hashes are retained with their individual records. Original statistical providers should be acknowledged alongside this collection: U.S. Census Bureau, Post-Secondary Employment Outcomes (PSEO); UK Department for Education, Longitudinal Education Outcomes (LEO); Korean Educational Development Institute / Korean Educational Statistics Service (KEDI/KESS); and the Chinese institutional reports identified in the source register. The US records distinguish analysis vintages from later coverage/licence records; the UK register identifies the exact retained release; the Korean register identifies the school and school-by-department series for 2020–2024; and the Chinese register preserves individual report and page locators. Consult the source registers rather than treating collection file counts as observation counts. DATA_USE_AND_REUSE_STATEMENT.txt, SOURCE_ATTRIBUTION_AND_NOTICES.txt and LICENSES_AND_SCOPE.txt explain provider acknowledgements and component-specific rights. The US public PSEO Flows catalogue identifies CC0 1.0; covered UK public statistical information carries the Open Government Licence v3.0 acknowledgement. Korean numeric-series reuse records are distinguished from general report, text and image policies. Chinese numerical extracts retain their individual source attributions. A licence for author-controlled contributions does not replace provider conditions or third-party software notices, and access to this collection does not grant rights over complete source publications. Cite the dataset using the DOI and version displayed on the published Zenodo record, together with the associated article, when available, when discussing its methods or scientific claims. CITATION.cff supplies the archive title and author order. Correspondence: Mengdi Yao ([e-mail]). Data and code contact: Yuchen Wang ([e-mail]).