AI Implementation Credibility: Data and Audit Package for the Corporate AI Transition
Description
This dataset supports the article “From AI Visibility to Organizational State: Implementation Credibility in the Corporate AI Transition.” The package contains the cleaned firm-year research data, implementation-credibility audit files, evidence-tier classifications, source-verification ledgers, falsification-oriented source-audit records, temporally lagged capability measures, Model B-T outputs, Model V cross-validation results, Model C downstream-validation results, state-transition diagnostics, variable dictionaries, claim-to-evidence maps, and reproduction scripts. The empirical design combines a longitudinal SEC-based firm-year panel with a 500 firm-year implementation audit. Public AI visibility is treated as an observable information signal, while implementation credibility is inferred from organizational capability, complementary assets, and independent implementation traces. The package documents the full evidence lineage used to distinguish symbolic high-gap, mixed partial backing, and substantively backed implementation classifications. The package is intended to support transparency, replication, and editorial or reviewer audit. Raw full-text SEC filings are not redistributed; public-source references and retrieval information are provided where applicable. Some source materials remain subject to the terms of their original providers.
Files
Steps to reproduce
1. Download and extract the complete data and audit package. 2. Review `00_package_documentation/README_IJIM_DATA_AUDIT_PACKAGE.md` for the package structure, sample definitions, evidence tiers, and interpretation boundaries. 3. Install the Python dependencies listed in `10_reproducibility/requirements.txt`. 4. Reproduce the central temporal implementation-credibility models by running: `python 10_reproducibility/reproduce_model_B_T.py` 5. Reproduce the grouped out-of-sample incremental-validity analyses by running: `python 10_reproducibility/reproduce_model_V.py` 6. Reproduce the complementary downstream-validation models by running: `python 10_reproducibility/reproduce_model_C.py` 7. Compare the generated `REPRODUCED_*` files with the supplied model-output CSV files. 8. Use `CLAIM_TO_EVIDENCE_MAP_CURRENT.csv`, `CLAIM_BOUNDARIES_CURRENT.csv`, and `MODEL_REGISTRY_CURRENT.csv` to trace manuscript claims to the corresponding data, estimates, and permitted inference. The supplied scripts reproduce the central Model B-T and Model V results. Model C coefficients and HC3 inference reproduce directly; small differences in firm-clustered covariance estimates may occur across statsmodels versions because of finite-sample corrections, as documented in `REPRODUCTION_VALIDATION.md`.