Smokescreen: A Python package for data vector blinding and encryption in cosmological analyses

Abstract: Smokescreen is an open-source Python library for data-vector concealment (blinding) in cosmological analyses. Data-vector blinding works by applying cosmology-dependent shifts to the observed data vector, moving it away from the true cosmological signal without affecting its statistical properties, so that analysts cannot infer the true result until the analysis is frozen and the blinding is lifted. The package computes these shifts using Firecrown likelihoods applied to data vectors stored in the SACC format, ensuring that the theoretical model used for blinding is identical to that used for inference whilst remaining agnostic to the specific observable being blinded. To prevent accidental unblinding, the original SACC file, containing the true cosmology, is encrypted. Although developed for the Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST), Smokescreen is applicable to any experiment using Firecrown likelihoods and the SACC data format.

• The paper introduces Smokescreen, a Python package that automates cosmological data vector blinding and encryption to mitigate experimenter bias.
• It integrates established frameworks like Firecrown and SACC to ensure that statistical inference and goodness-of-fit tests remain robust despite data concealment.
• Validation on LSST-like simulations shows that the blinding strategy maintains physical plausibility with negligible shifts in posterior Bayesian log-evidence.

Smokescreen: A Standard Python Infrastructure for Data Vector Blinding and Encryption in Cosmological Analyses

Motivation and Context

The integrity of statistical cosmological inference is threatened by experimenter and confirmation biases, especially as precision escalates in forthcoming analyses of data from projects like LSST. Unlike experimental disciplines with more direct double-blind methodologies, cosmology is burdened by the uniqueness of observable data—global correlation patterns in a single universe—which precludes simple blinding via signal region masking. The risk of analysts tuning pipelines to expectations or external results calls for community-wide, transparent, and reproducible blinding procedures at the data vector level.

Smokescreen directly addresses these challenges, providing an open-source, Python-based infrastructure that automates data-vector concealment using cosmology-dependent shifts, robust file encryption, and seamless integration with established survey analysis pipelines via SACC and Firecrown conventions (2604.18111).

Methodology and Core Algorithm

At its foundation, Smokescreen implements the data vector blinding strategy as formalized by Muir et al. (2020), which satisfies the requirements of masking the true cosmological result while ensuring that posterior inference remains physically plausible and validation tests (e.g., systematics null and goodness-of-fit) are unimpeded. The core algorithm follows a five-step protocol:

1. Parameter Shift Sampling: For each selected cosmological parameter, shifts are drawn from user-defined deterministic, uniform, or Gaussian distributions.
2. Blinded Cosmology Construction: The blinded cosmology vector $\Omega_{\text{blind}}$ is derived by adding the sampled parameter shifts to the reference cosmology.
3. Data Vector Prediction: The theory vector for both reference and blinded cosmologies is obtained via an external likelihood framework (Firecrown) with systematics parameters ($S_{\text{ref}}$) fixed.
4. Blinding Factor Calculation: The difference between predictions (additive or multiplicative, as appropriate) defines the concealing transformation.
5. Application and Encryption: The measured data vector is shifted accordingly, creating the blinded data product. The original data vector is encrypted using Fernet (AES-128-CBC+HMAC-SHA256), and only the relevant CLI-configured parties possess the recoverable key, securely separating blinding and analysis teams.

This approach ensures that the analyst cannot infer the cosmology until the encrypted original is restored, while the statistical structure and likelihood interface remain untouched, preserving downstream validation capability.

Software Architecture and Integration

Smokescreen’s design adheres strictly to separation of concerns, with distinct modules for parameter perturbation, likelihood-based theory evaluation, encryption and decryption, and CLI orchestration. Its central abstraction, ConcealDataVector, coordinates these roles, facilitating modular testing and rapid modification of parameter shift strategies.

Compatibility with Firecrown enables seamless integration with established cosmological analysis platforms (CosmoSIS, Cobaya, NumCosmos), while SACC ensures robust data handling and metadata preservation. Blinding configurations are both CLI and YAML-configurable, and all outputs (FITS/HDF5 SACC) are validated for floating-point consistency, parameter integrity, and completeness. The inclusion of audit metadata (seed, user, timestamp) supports full reproducibility and transparency.

The unit test suite achieves 94% code coverage, with rigorous correctness checks across physical, configuration, and I/O integrity regimes.

Empirical Validation and Results

Smokescreen was validated on simulated LSST Y1 3x2pt multi-probe data vectors, including cosmic shear, galaxy clustering, and galaxy-galaxy lensing across five tomographic redshift bins. Deterministic shifts were applied to pivotal parameters ($A_s$, $w$), producing “Blind A” and “Blind B” realizations.

Posterior analyses using CosmoSIS and the identical Firecrown likelihood demonstrate that the maxima of the blinded posteriors move in the prescribed direction, with negligible change in Bayesian log-evidence ($|\Delta \log Z| \leq 1$). Goodness-of-fit statistics remain invariant, confirming that data validation and null testing are unhampered by the concealment shift (satisfying criteria from Muir et al., 2020). Small residual shifts in correlated parameters (e.g., $\Omega_c$) reflect expected degeneracies and confirm the physical plausibility of the procedure.

Notably, the blinding and inference models are mathematically identical by construction, obviating past sources of implementation-induced bias.

Security, Practicality, and Reproducibility

Security is ensured by using established cryptographic primitives, with each blinding instance utilizing a newly generated Fernet key and audit-trailed configuration, precluding accidental or unauthorized unblinding. CLI-facing commands abstract the procedural complexity for end-users, and configuration options are both explicit and extensible.

Smokescreen’s reproducibility infrastructure—seeded randomness, parameter validation, audit metadata—aligns with modern open science standards, enabling post-unblinding cross-checks by independent reviewers.

Impact, Broader Applicability, and Future Prospects

Smokescreen fills a structural gap in the cosmology software ecosystem, standardizing data vector blinding for the LSST DESC pipeline and already seeing adoption beyond its initial scope (e.g., KIDS 6x2pt legacy analyses). Its modular design and reliance on open standards (SACC, Firecrown) facilitate future expansion to additional Stage-IV dark energy surveys and support cross-collaboration harmonization of blinding practices.

This infrastructure mitigates experimenter bias and improves the credibility of highly sensitive cosmological inferences. As multi-probe analyses and joint likelihood pipelines become ubiquitous, tools like Smokescreen are likely to become indispensable for the community, potentially driving convergence toward universal blinding protocols and enhanced reproducibility mandates.

Conclusion

Smokescreen constitutes a robust, standardized infrastructure for data vector blinding and encryption, directly enabling bias mitigation in modern cosmological inference pipelines. By automating, standardizing, and securing the blinding protocol, Smokescreen substantially augments the reproducibility, transparency, and reliability of large-scale analyses—a critical advance as datasets and model complexities accelerate in the coming survey era (2604.18111).