T7-replication-package-v2.0

# Replication package: Refactoring a Production Legacy System with LLM AgentsVersion 1.0 - snapshot built 2026-09-28T18:18:54.Published with the paper submitted to *Information and Software Technology*.## What is in this archive| Path | Content ||---|---|| `tasks/tasks.json` | 54 frozen method-level refactoring tasks (v1.0) || `tasks/structure_tier.csv` | measured structural features and tiers per task || `prompts/` | frozen prompt texts, tool schema, and their SHA-256 || `src/` | agent harness, tools, verification and analysis code || `env/` | Dockerfile, PMD ruleset, environment fingerprint || `logs/` | JSONL event logs for every run (de-identified) || `results/` | dual-gate outcome dataset and all analysis outputs || `c1_baselines/` | human-baseline diffstat manifest || `protocol/` | experimental protocol and pre-registration text || `MANIFEST.csv`, `CHECKSUMS.sha256` | every file with size and SHA-256 |## Snapshot| Quantity | Value ||---|---|| runs in `results/p4_runs.csv` | 137 || runs by model tag | {'flash': 82, 'pro': 55} || dual-gate passes | 29 || JSONL run logs | 412 || patch metadata files | 134 || files in manifest | 615 |## Outcome definitionA run counts as a pass only if both gates succeed: `mvn compile` succeeds(`compile_ok`) **and** the PMD complexity thresholds are met(`complexity_gate == "pass"`). The derived column is `dual_pass`.The `verify` column of the older `results/p4_results.csv` records thecomplexity gate only and must not be used as the outcome.## How to reproduce```bash# 1. environment: JDK 8, Maven 3.9.x, PMD 7.28, Python 3.8pip install -r env/requirements.txt# 2. point the harness at your own Java 8 + Spring Boot + MyBatis checkoutcp config.example.yaml config.yaml # set workspace + baseline commit# 3. run the matrix (task x condition x repetition)python src/harness/run_p4_v2.py --tag flash --model deepseek-chat --tasks all# 4. rebuild the dataset from the logs and re-run the analysispython src/harness/build_dataset.pypython src/analysis/structure_tier.py # measured structure tierspython src/analysis/run_behavior_check.py # behaviour-preservation proxypython src/analysis/stats.py # GEE, McNemar, effect sizes```The studied source code is proprietary and is not included; the harnessonly needs a git checkout, `mvn compile`, and PMD, so any comparableJava 8 / Spring Boot / MyBatis system can be substituted. See`DATA_AVAILABILITY.md` for what is released and what is withheld.## Licence and citationSource code: MIT. Derived data and logs: CC BY 4.0. The proprietarysystem source and the human-baseline diffs are not licensed forredistribution and are not included. See `LICENSE` and `CITATION.cff`.## De-identificationCompany, repository, package, identifier prefixes, local paths andcommit hashes were replaced mechanically; source quotations in the logswere removed. Substitution rules are in `ANONYMIZATION.md`; the originalvalues are not distributed.