Jev at the Agent Authorization Boundary: Evidence, Data, and Reproduction Code

Jev at the Agent Authorization Boundary: Evidence, Data, and Reproduction Code Version 1.2.0. All-versions DOI: https://doi.org/10.5281/zenodo.22860054 This version adds post-review diagnostics for manuscript version 2. Read REVIEW-DIAGNOSTICS.md. It keeps the bootstrap correction from version 1.1.0 (https://doi.org/10.5281/zenodo.22863437; read STATISTICAL-CORRECTION.md) and the original observations of version 1.0.0 (https://doi.org/10.5281/zenodo.22860055). Original observations, labels, selected thresholds and freeze files are unchanged. Historical analyze.py files remain as provenance and contain the superseded bootstrap defect; do not use their analyze command as the current entry point. Reproduction With Python 3.12 or newer, install requirements.txt in an isolated environment. Run `python verify_evidence.py`, then `python reproduce_evidence.py`. No experimental model requests or credentials are needed. The wrapper regenerates historical results, checks the corpus and frozen development thresholds, runs corrected analyses, independently validates every paired Brier interval, and regenerates and cross-checks the post-review diagnostics. All added diagnostics are descriptive and do not select new gates from test outcomes. The deposit excludes the manuscript. It contains a historical 150-fixture audit and two separate controlled sessions on 108 development and 270 test cases. The same test families are reused across sessions. Policies and reference labels are authored synthetic targets, not independently adjudicated policy truth. Static proposals do not exercise an enforcing executor. Proprietary RLCD training is not experimentally isolated. Provenance and licenses SHA256SUMS.json binds the original distributed file inventory before replay. Replay can create Python caches and regenerate derived outputs; use a fresh extracted copy for archive-integrity verification. deposit-provenance.json maps copied files to the accompanying source-release integrity records. Source-provenance includes records for manuscript files deliberately excluded here. Raw observations and original freeze checks remain verifiable. Original data and documentation use CC BY 4.0; original Python code uses MIT. See LICENSES.md. Third-party rights are retained. AI assistance supported implementation, synthetic construction, analysis, and documentation, including Anthropic-based assistance for the version 1.2.0 diagnostics. No additional human annotation was performed.