Replication Package for Agentic Cryptographic Debt: Repository-Level Measurement of Post-Quantum Migration Regression Under Autonomous AI Software Development

Replication package for the manuscript "Agentic Cryptographic Debt: Repository-Level Measurement of Post-Quantum Migration Regression Under Autonomous AI Software Development" (Robert Campbell, submitted to MDPI Computers, 2026). The study measures what two open-weight coding models produced when asked to perform a repository-level migration from RSA-based JWT signing to ML-DSA (FIPS 204) across four conditions and a nested ablation, scored by static analysis and behavioural execution. The package is indexed so that every quantitative result and quoted excerpt in the paper can be traced to a run identifier. MANIFEST.md lists every file with its size and SHA-256. RUN_INDEX.md gives one row per scored run: repository, feasibility, model and checkpoint where recorded, condition, prompt variant, run identifier, the paths of the run record and transcript, the assembled-prompt hash where recorded, the per-layer findings, and the merged verdict where the record carries one. Both are generated by build_manifest.py, which also audits the package against its traceability requirements and reports what is present, partial, or absent. Contents: the harness and bounded agent loop; the scorer, all three layers as implemented, with its validation fixtures and the scoring specification defining every finding code; the task specifications for all four primary conditions and the nested ablation; the seeded classical baselines; the run record and transcript for each of the 56 scored runs, including build output and, where captured, test output; the seven excluded records in a separate directory so the exclusions can be inspected; the model-provenance bundle with manifest and weights digests, quantization, and context length; the feasibility-audit evidence bundle with per-library install digests, registry contents, and verbatim exceptions; the context-sensitivity ablation responses with their re-derivation script; environment lockfiles and the pinned toolchain versions; the apparatus-defect correction chronology; and the verification scripts verify_allocation.py, verify_apparatus_currency.py, rescore_record.py, rescore_from_raw.py, and count_green_states.py. A README with commands to reproduce a single run, a LICENSE, and a third-party licensing notice are included. Coverage is not uniform and is stated rather than left to be discovered: some fields are absent from records executed before the corresponding apparatus fix, and the feedback and agentic runners write their own record structures. The transcripts record the path and byte size of each file the agent wrote but not its contents. Pre- and post-run git-tree hashes were not recorded by the harness. The direct-question probes of the paper's Section 6.3 are outside the scored dataset and are not archived as run records. The paper's Data Availability Statement gives the per-field counts. The two corpus repositories are third-party open-source projects, cited by commit hash in the paper, and are not redistributed here. Licensed under Creative Commons Attribution 4.0 International; no patent licence, express or implied, is granted by this publication. Corresponds to git tag v1.0.0.