Launder-Bench

Can repository-level agentic code laundering be detected?

Launder-Bench evaluates whether repository provenance remains recoverable after a coding agent reconstructs a real repository from behavioral specifications with the source hidden, while preserving its tested behavior. Every released rewrite passed execution-based acceptance gates; labels follow provenance paths, not surface similarity.

• Code and full data pipeline: github.com/HuieL/Launder-Bench (source and reference collection, agentic rewriting, behavioral verification, task construction, detector evaluation).
• Status: this card describes the release; the data files will be added to this repository.

Composition

Component	Size	Description
Source repositories	282	pinned real projects: 199 Python, 83 Java (URL, commit, snapshot date)
Same-domain references	1,244	measured independent repositories per source, TF-IDF similarity < 0.35 (median 0.076)
Launder-Pair	1,128	4 verified rewrites per source, each with its reference neighborhood
Launder-Attr	229	one query per source with 5 measured references, over 20 fixed candidates
Control tier	150 / 30	random, unselected Pair / Attr instances over 3 axes

Rewriting axes: py→js, py→py, java→java, java→js. The release is the full verified set; no detector-informed or difficulty-based selection defines it.

Tasks

Launder-Pair — pairwise provenance detection. Given repositories (R_q, R_c), decide whether they share a source ancestor.

Relation	Label	Meaning
source–rewrite	+	direct provenance
rewrite–rewrite	+	shared provenance through the same source
rewrite–independent	−	matched purpose, no provenance path (central hard negative)
rewrite–unrelated	−	neither matched purpose nor provenance

Launder-Attr — source attribution. Rank a fixed set of 20 distinct candidates that contains the true source exactly once (its 5 measured references are the hard alternatives; language-matched distractors complete the set). A detector may abstain.

Metrics. AUROC and best F1 for pairwise discrimination; band-calibrated TPR/FPR; tie-aware Rank@1 and calibrated attribution. The source-specific reference band is β_d(s) = max over the source's references N of d(R_s, N); calibrated attribution credits the true source only when it is ranked first and its score exceeds that band. A source-absent diagnostic removes the true source from the candidate set.

Construction

Rewrites are produced by Codex CLI 0.144.5 with GPT-5.6-Sol (GLM-5.3 and DeepSeek-V4.1-Flash in the backbone ablation). Each stage is a separate headless session: a decompose session writes a bottom-up unit manifest, the source is hidden, each unit is rewritten in isolation from its own specification and its dependencies' interfaces, and a compose session assembles the units and iterates against the behavioral oracle until every golden case passes. A second cycle re-decomposes the composed repository, and a critic removes surviving source-chosen tokens. Three recipes are released: full contract, external vocabulary and critic.

A rewrite is accepted only if the source harness yields at least 5 successful cases, the rewrite reproduces at least half of the successful golden outputs exactly with overall conformance ≥ 0.5, its conformance adapter stays thin (≤ half of the artifact's tokens), and its size is judged against min(original tokens, 11,000).

Contents of a release instance

• the rewritten repository and its conformance adapter (excluded from scoring);
• the source's conformance cases and golden traces;
• the pinned source manifest and the reference neighborhood;
• the Pair relations or the fixed Attr candidate list, with axis and recipe.

Limitations and responsible use

Behavioral validity is oracle-covered rather than whole-program equivalence. Reference repositories are empirical controls, not certificates of legal or historical independence. Source repositories are third-party projects under their own licenses; the benchmark records technical derivation and does not determine legal ownership or license compliance.

Citation

@inproceedings{launderbench2027,
  title     = {Launder-Bench: Can Repository-Level Agentic Code Laundering Be Detected?},
  author    = {Anonymous},
  booktitle = {Under review},
  year      = {2026}
}