consistency: report adds Code & Data Availability + 4 cross-artefact fixes

Four small repo-wide consistency fixes plus a new "Code and Data
Availability" section in the report so reviewers can locate every
artefact described in the paper.

- cleantest_agent/__init__.py: replace the hard-coded
  __version__ = '0.1.0' (drifted past 0.1.1 on the last release)
  with importlib.metadata.version('cleantest-agent'). pyproject.toml
  is now the single source of truth for the version, and python-side
  __version__ tracks pip's view of the installed distribution.
- docs/skill-distribution-guide.md: replace the version literal in
  the import-and-print walkthrough with X.Y.Z + a note that the
  string is resolved at import time, not embedded.
- report/references.bib: add a self-cite @misc{yang2026cleantestagent}
  pointing at github.com and pypi.org so the BibTeX corpus matches the
  citation stanza already published in README.md.
- report/main.tex:
  * \authornote on the title page links the GitHub + PyPI URLs and
    forwards the reader to the new section. (\thanks on \title
    blew up acmart's metadata stack; \authornote is the
    template-canonical equivalent.)
  * Introduction outline now ends with one sentence pointing at the
    availability section, then a paragraph that names every artefact
    (package, skills, tests, CI matrix, CD pipeline, Filter 3
    metrics, LaTeX sources) and cites the self-reference.
  * New Section 10 "Code and Data Availability" (§10 in the new
    numbering; Conclusion is §9) carries a 14-row table mapping every
    empirical claim to its reproducible source, plus three
    paragraphs on reproduction, supply-chain integrity (sigstore),
    and contribution channels.
- report/main.pdf: rebuilt locally with TeX Live 2026; 67 pp. -> 68 pp.,
  bibliography now has 40 entries (was 39), all citations resolved,
  no missing references.

Author: jimmy0717

cleantest_agent/__init__.py

@@ -9,7 +9,21 @@
     )
 """
 
-__version__ = "0.1.0"
+# Single source of truth for the version is `pyproject.toml`;
+# `__version__` is resolved at import time from the installed
+# distribution metadata so the two cannot drift out of sync.
+# Falls back to "0.0.0+unknown" only when the package is imported
+# directly from a source tree that has never been `pip install`-ed
+# (e.g. by running tests from a freshly cloned repo without pip);
+# all CI / CD / PyPI / sigstore paths always go through an install.
+try:
+    from importlib.metadata import PackageNotFoundError, version as _pkg_version
+    try:
+        __version__ = _pkg_version("cleantest-agent")
+    except PackageNotFoundError:
+        __version__ = "0.0.0+unknown"
+except ImportError:  # pragma: no cover - importlib.metadata is stdlib on 3.8+
+    __version__ = "0.0.0+unknown"
 
 from cleantest_agent.data_loader import load_csv, save_csv
 from cleantest_agent.report_generator import NoiseReport

docs/skill-distribution-guide.md

@@ -54,7 +54,9 @@ Verify:
 
 ```bash
 python -c "import cleantest_agent; print(cleantest_agent.__version__)"
-# 0.1.0
+# X.Y.Z   (e.g. 0.1.1; the version is resolved at import time from
+#          the installed distribution metadata, so it always tracks
+#          whatever pip installed)
 
 cleantest --help
 # usage: cleantest [-h] --input_csv INPUT_CSV --output_dir OUTPUT_DIR ...

report/main.pdf

@@ -106,6 +106,14 @@
 \title[CleanTest-Agent]{CleanTest-Agent: A Multi-Agent Skill-Orchestrated System for Unit Test Training Data Quality Assurance}
 
 \author{Yong Yang}
+\authornote{Source code, datasets, the trained Filter 3 checkpoint
+  metadata, and a reproducible end-to-end notebook are released
+  under the MIT license at
+  \url{https://github.com/jimmy0717/cleantest-agent}; a
+  pip-installable distribution is at
+  \url{https://pypi.org/project/cleantest-agent/}. See
+  Section~\ref{sec:availability} for the full availability
+  statement.}
 \affiliation{%
   \institution{Beihang University}
   \department{School of Software}
@@ -189,7 +197,15 @@ \section{Introduction}
     \item \textbf{Empirical evaluation}: We compare four approaches (rule-based, zero-shot LLM, few-shot LLM, hybrid) and demonstrate that our hybrid approach achieves the best F1 (0.965) while the pure LLM approach achieves only 0.387, validated with real DeepSeek-V4-Flash API calls.
 \end{enumerate}
 
-The remainder of this paper is organized as follows. Section~\ref{sec:background} reviews related work on unit test generation, data quality, LLM-based code analysis, and agent skill architectures. Section~\ref{sec:requirements} presents a requirements analysis including stakeholder analysis, use cases, functional and non-functional requirements, and a traceability matrix. Section~\ref{sec:model-driven} defines the model-driven approach and contrasts it with the pure-LLM approach. Section~\ref{sec:design} details the system design including architecture, filter specifications, and the pipeline algorithm. Section~\ref{sec:implementation} covers the implementation with technology choices, design decisions, and testing strategy. Section~\ref{sec:evaluation} presents the experimental evaluation through four research questions, including a case study on real samples and an ablation study. Section~\ref{sec:discussion} discusses findings, lessons learned, connections to course topics, and threats to validity. Section~\ref{sec:conclusion} concludes the paper and outlines future work directions. Appendices provide LLM prompt templates, CI/CD configuration, the full noise report, noise type examples with real Java code, a complete API reference, and skill directory structure documentation.
+The remainder of this paper is organized as follows. Section~\ref{sec:background} reviews related work on unit test generation, data quality, LLM-based code analysis, and agent skill architectures. Section~\ref{sec:requirements} presents a requirements analysis including stakeholder analysis, use cases, functional and non-functional requirements, and a traceability matrix. Section~\ref{sec:model-driven} defines the model-driven approach and contrasts it with the pure-LLM approach. Section~\ref{sec:design} details the system design including architecture, filter specifications, and the pipeline algorithm. Section~\ref{sec:implementation} covers the implementation with technology choices, design decisions, and testing strategy. Section~\ref{sec:evaluation} presents the experimental evaluation through four research questions, including a case study on real samples and an ablation study. Section~\ref{sec:discussion} discusses findings, lessons learned, connections to course topics, and threats to validity. Section~\ref{sec:conclusion} concludes the paper and outlines future work directions. Section~\ref{sec:availability} states code, data, and reproducibility links. Appendices provide LLM prompt templates, CI/CD configuration, the full noise report, noise type examples with real Java code, a complete API reference, and skill directory structure documentation.
+
+All artefacts described in this paper --- including the
+Python package \texttt{cleantest-agent}, the four SKILL.md skill
+bundles, the 36-case pytest suite, the GitHub Actions CI matrix, the
+tag-driven PyPI publish workflow with sigstore signing, the curated
+Filter 3 model-mode metrics, and the LaTeX sources of this paper
+---  are released under the MIT license and available at
+\url{https://github.com/jimmy0717/cleantest-agent}~\cite{yang2026cleantestagent}.
 
 \subsection{Scope and Context}
 
@@ -3077,6 +3093,78 @@ \subsection{Final Reflections}
 
 The skill-based architecture also illustrates how modern software design patterns adapt to new paradigms: the \texttt{SKILL.md} protocol is effectively an interface contract between the skill author and the AI coding assistant runtime --- a natural evolution of API design for the age of natural-language-driven software interaction.
 
+% ============================================================================
+\section{Code and Data Availability}
+\label{sec:availability}
+% ============================================================================
+
+In line with the ACM Artifact Review and Badging policy, all
+artefacts described in this paper are publicly available under the
+MIT license. Table~\ref{tab:availability} maps every empirical claim
+in the paper to its reproducible source.
+
+\begin{table}[htbp]
+\centering
+\small
+\caption{Code and data availability map.}
+\label{tab:availability}
+\begin{tabularx}{\linewidth}{@{}lY@{}}
+\toprule
+\textbf{Artefact} & \textbf{Location} \\
+\midrule
+Source repository (MIT)            & \url{https://github.com/jimmy0717/cleantest-agent} \\
+Python package (pip-installable)   & \url{https://pypi.org/project/cleantest-agent/} \\
+Tagged release (this paper)        & \url{https://github.com/jimmy0717/cleantest-agent/releases/tag/v0.1.1} \\
+This paper (PDF, signed asset)     & \url{https://github.com/jimmy0717/cleantest-agent/releases/download/v0.1.0/main.pdf} \\
+SKILL.md skill bundles (4)         & \texttt{skills/} in the repository \\
+Filter 3 training notebook         & \texttt{experiments/main-final.ipynb} \\
+Filter 3 metrics (held-out test)   & \texttt{experiments/results/coverage\_run/test\_metrics.json} \\
+Bundled 5{,}000-row sample         & \texttt{data/sample\_5000.csv} (Methods2Test, MIT) \\
+Annotation dictionary (21{,}954 patterns) & \texttt{cleantest\_agent/data/noise\_modifier\_fm.txt} \\
+500-sample evaluation labels       & \texttt{experiments/results/labeled\_samples.csv} \\
+DeepSeek baseline runner           & \texttt{experiments/run\_baselines.py} \\
+CI workflow (Python 3.10/3.11/3.12) & \texttt{.github/workflows/ci.yml} \\
+CD workflow (PyPI + sigstore)      & \texttt{.github/workflows/publish.yml} \\
+Upstream LessIsMore-FSE2025 data   & \url{https://doi.org/10.5281/zenodo.15347368} \\
+\bottomrule
+\end{tabularx}
+\end{table}
+
+\textbf{Reproducing the headline results.}
+The 593{,}953-sample full-pipeline run reported in
+Section~\ref{sec:evaluation} (RQ1) is reproducible by cloning the
+repository, exporting Methods2Test via the LessIsMore-FSE2025
+replication package, and invoking
+\texttt{cleantest --input\_csv all\_train.csv --output\_dir out/}.
+The 500-sample stratified subset evaluation (RQ2--RQ4) is
+reproducible via
+\texttt{python experiments/run\_baselines.py}, which issues real
+DeepSeek-V4-Flash API calls; the per-sample predictions used for
+the F1\,=\,0.965 hybrid result and the LLM zero-shot / few-shot
+baselines are checked into
+\texttt{experiments/results/labeled\_samples.csv}. The Filter 3
+model-mode held-out MAE\,=\,0.0309 result is reproducible from
+\texttt{experiments/main-final.ipynb} on a single A800 80\,GB; the
+metrics archived under
+\texttt{experiments/results/coverage\_run/} are the verbatim
+artefacts that produced the numbers in
+Section~\ref{sec:evaluation} (RQ3 model mode).
+
+\textbf{Supply-chain integrity.}
+PyPI distributions for v0.1.1 onward are uploaded via OIDC Trusted
+Publisher (no long-lived API token) and signed with sigstore;
+the signed wheel and source bundles are attached as release assets
+on GitHub. End users can verify provenance with the
+\texttt{sigstore verify identity} command documented in
+\texttt{docs/PYPI-PUBLISHING.md}.
+
+\textbf{Issue tracking and contribution.}
+The repository ships with structured GitHub issue templates (bug,
+feature, question), a pull-request template, the Contributor
+Covenant 2.1 code of conduct, and a development-workflow
+\texttt{CONTRIBUTING.md}. Bug reports, feature requests, and pull
+requests are welcomed at the URLs above.
+
 % ============================================================================
 % Acknowledgements (ACM-recommended `acks` environment, not `\section{...}`).
 % ============================================================================

report/main.tex

@@ -314,3 +314,15 @@ @book{meszaros2007xunit
         modalities of automated test design (Chapters 11--12); referenced
         by Filter 2's reflection rule checklist (R2, R3) in this paper.}
 }
+@misc{yang2026cleantestagent,
+  title        = {{CleanTest-Agent}: A Multi-Agent Skill-Orchestrated System
+                  for Unit Test Training Data Quality Assurance},
+  author       = {Yang, Yong},
+  year         = {2026},
+  howpublished = {\url{https://github.com/jimmy0717/cleantest-agent}},
+  note         = {Source code, datasets, and a reproducible Filter 3
+                  training notebook are released under the MIT license at
+                  \url{https://github.com/jimmy0717/cleantest-agent}; the
+                  pip-installable package is at
+                  \url{https://pypi.org/project/cleantest-agent/}.}
+}