RECAST: Extending the Impact of Existing Analyses

Searches for new physics by experimental collaborations represent a significant investment in time and resources. Often these searches are sensitive to a broader class of models than they were originally designed to test. We aim to extend the impact of existing searches through a technique we call ‘recasting’. After considering several examples, which illustrate the issues and subtleties involved, we present RECAST, a framework designed to facilitate the usage of this technique.

💡 Research Summary

The paper addresses a fundamental bottleneck in high‑energy physics: the difficulty of re‑using existing experimental searches to test new theoretical models that were not originally considered. While large collaborations invest substantial resources in designing, executing, and publishing analyses, the raw data and detailed analysis code are typically kept within the collaboration, making it cumbersome for external theorists to reinterpret the results. To overcome this, the authors introduce the concept of “recasting,” which consists of keeping the original selection criteria, background estimates, and statistical treatment intact, and simply substituting a new signal model’s simulated events into the established workflow. The central contribution of the work is RECAST, a software framework that operationalizes recasting in a systematic, secure, and reproducible manner.

RECAST is organized into three layers. The front‑end is a web‑based portal where users upload a model description (e.g., a UFO file), a set of Monte‑Carlo events, and any auxiliary information required for the reinterpretation. The portal automatically validates the submission, assigns a unique request identifier, and forwards the payload to the back‑end. The back‑end runs inside an authenticated, containerized environment (Docker or Singularity) on the collaboration’s computing cluster or on a cloud service. Within this sandbox, the uploaded events are processed through the exact same ROOT‑based analysis chain that was used for the original search: the same trigger emulation, object reconstruction, event selection cuts, and histogram filling scripts are executed without modification. Detector effects are reproduced either with a full GEANT4 simulation (when resources allow) or with a fast‑simulation tool such as Delphes, calibrated using efficiency maps and smearing functions supplied by the original analysis.

A key technical challenge is the treatment of systematic uncertainties. RECAST solves this by importing the original analysis’s nuisance‑parameter model, including priors, covariance matrices, and any shape systematics. These are attached to the new signal likelihood, ensuring that the statistical inference (e.g., CLs limits or profile‑likelihood ratios) remains fully consistent with the original publication. The framework also records the full chain of software versions, configuration files, and random seeds, guaranteeing reproducibility.

The authors demonstrate the utility of RECAST through three detailed case studies. First, they recast the 125 GeV Higgs boson search to constrain models where the Higgs decays invisibly, achieving limits comparable to dedicated analyses but with a turnaround time of only a few weeks. Second, they reinterpret a supersymmetry (SUSY) search by varying the mass hierarchy of neutralinos and charginos, thereby opening up previously unexplored regions of parameter space. Third, they apply the framework to a long‑lived particle (LLP) search, showing that the original trigger and detector‑subsystem efficiencies can be reused to test new LLP lifetimes and decay topologies. In each example, RECAST reproduces the original background expectations and systematic error budget, while delivering new signal‑specific results with minimal additional computational cost.

Security and access control are integral to the design. All back‑end jobs run under a least‑privilege user account, and authentication is handled via OAuth‑style tokens issued by the collaboration’s identity provider. This prevents unauthorized access to internal data while still allowing the necessary information (e.g., background histograms, systematic covariance) to be shared with the requestor.

In the discussion, the authors argue that RECAST can become a cornerstone of a “living analysis” ecosystem, where published results remain dynamically extensible as theory evolves. By providing a standardized API and encouraging collaborations to expose analysis metadata (selection cuts, systematic models, likelihood functions), the framework can be adopted across experiments such as ATLAS, CMS, Belle II, and future facilities like DUNE. The paper concludes with a roadmap that includes expanding the library of supported fast‑simulation back‑ends, integrating machine‑learning‑based object classifiers, and fostering an open‑source community to maintain and extend the RECAST code base. In essence, RECAST transforms a static publication into an interactive tool, dramatically amplifying the scientific return on the original experimental investment.