Benchmark Release · v1.0.0 · Live Package · 1.2.0 · Clean DOI · 10.5281/zenodo.19475781

Dormant Behavior Audit

A benchmark-first release for discovering, validating, and reporting latent model behavior with reproducible artifacts, explicit controls, and claim-level evidence.

Open Release PyPI Package Zenodo DOI Download Report Reference Bundle GitHub Repo Reviewer Quickstart
0 / 490 competitor false positives in the checked-in reference case
34.1% model-2 pooled rate on top Alibaba-family triggers
13.5%–22.0% model-3 pooled top-5 band, active but materially weaker
37.3% vs 3.3% `马云` split between model-2 and model-3

Flagship Reference Case

Finding the Alibaba Cloud Backdoor

The public release centers on a reproducible reference case that has been normalized into a benchmark bundle, paired with appendices, validation records, and release-ready evidence artifacts.

Slow terminal tour of the benchmark release path

What Ships

Benchmark, evidence, and release assets

  • Benchmark charter, governance, and launch docs
  • Normalized dormant puzzle reference bundle
  • Public stateful multi-turn candidate plus Qwen2 and Qwen2.5 clean-control lanes with repeat anchors
  • Release-state, claim-ledger, and artifact-hash guardrails
  • Reviewer quickstart, traceability matrix, and arXiv endorsement packet
  • Reproducibility and tightening bundles
  • Interpretation-aware hosted follow-up packets
  • Public collaboration and submission materials

Install

Package and CLI

pipx install dormant-behavior-audit
dba doctor
dba scoreboard
dba reviewer-packet --out-root reviewer_packet

The PyPI package currently ships the full research stack, so installation is heavier than a tiny utility CLI by design.

Reproduce

Fastest credible rerun path

python3 scripts/build_reviewer_packet.py --out-root reviewer_packet
python3 scripts/reproduce_submission.py --report-only \
  --out-root artifacts/reproduction/20260305_230206

Judge success with the reproduction report and claim-consistency report rather than exact JSON replay for stochastic hosted systems.

Collaborate

Best first ways to engage

  • Replicate a seeded task or control family
  • Package a method submission against the benchmark contract
  • Contribute calibration or mechanism-characterization tasks
  • Help validate portability, evidence packaging, or scoring surfaces
Read the collaboration brief

Cite

Use the tagged release and report

Cite the repository release materials from CITATION.cff and the reference report titled Finding the Alibaba Cloud Backdoor: A Reproducible Reference Case for Dormant Behavior Audit.

Version DOI: 10.5281/zenodo.19475781

Open citation metadata