---
title: 'GAMBIT Corpus: Global BSM Inference Framework'
url: https://www.emergentmind.com/topics/gambit-corpus
type: topic
---

# GAMBIT Corpus: Global BSM Inference Framework

GAMBIT Corpus denotes the body of software, model definitions, interfaces, and inference workflows centered on **GAMBIT**, the **Global and Modular BSM Inference Tool**, for statistically rigorous global fits of beyond-the-Standard-Model theories. In this usage, “corpus” refers not to a text collection but to an open-source, massively parallel ecosystem that combines heterogeneous observables and likelihoods from collider searches, dark matter, flavour physics, precision measurements, neutrino physics, and cosmology within a single inference pipeline [2105.03165]. The framework overview for the first public release described GAMBIT as a modular, model-aware system intended to make global fits safer, faster, reproducible, and more easily extendible than earlier model-specific tools [1705.07908].

## 1. Concept and motivation

The central motivation for the GAMBIT corpus is that BSM theories are constrained simultaneously by many complementary probes, and that robust assessment therefore requires a single composite likelihood rather than isolated limits. The original introduction to GAMBIT emphasized the need to combine direct searches for new particles at colliders, Higgs measurements, flavour observables, relic-density information, direct and indirect dark-matter searches, and electroweak precision observables into one consistent statistical analysis [1611.05065]. Later descriptions extended that motivation to neutrino and cosmological data, presenting global fits as the most accurate and statistically sound methods for evaluating which regions of parameter space survive, which are preferred, and where tensions arise [2105.03165].

Within this conception, the corpus is methodological as much as phenomenological. It exists to turn model testing into a unified inference problem. The axion study made this logic explicit: one should not simply overlay individual exclusion curves, but instead combine all available information in a unified likelihood-based framework that can compare models, extract parameters, and assess discovery potential [1710.11138]. A similar argument appears throughout the supersymmetry applications, where collider, dark-matter, flavour, and precision inputs constrain different mechanisms and different corners of parameter space [1805.06049].

## 2. Architectural principles

GAMBIT is organized around a modular architecture in which the basic computational units are **module functions** identified by **capabilities**, **types**, optional **dependencies**, optional **backend requirements**, and model restrictions [1705.07908]. This capability-based abstraction allows the framework to construct a directed acyclic graph of calculations at runtime, so that a likelihood term can request a quantity such as a mass, cross section, decay rate, or nuisance parameter without hard-coding where that quantity originates. The resulting dependency resolver activates compatible functions, resolves backend requirements, and orders calculations automatically [1705.07908].

A second foundational element is the **hierarchical model database**. Models are represented as named parameter sets, arranged in parent/child relationships through translation functions, with additional friend-model mappings where needed [1705.07908]. This makes it possible to relate low-parameter and high-parameter realizations of the same theory, to vary nuisance parameters jointly with BSM parameters, and to reuse calculations across related model families. The 2021 overview singled out this model database as an important mechanism for translating between different parameterizations, using the example of converting a low-parameter MSSM version into a more general one [2105.03165].

The backend system is equally central. GAMBIT can interface dynamically to external specialist codes through POSIX dynamic loading for C, C++, or Fortran backends, WSTP for Mathematica, and pybind11 for Python backends [2105.03165]. The framework paper further described additional machinery such as BOSS for runtime interaction with C++ classes and CBGB for generating frontend boilerplate for Fortran common blocks [1705.07908]. This design lets GAMBIT reuse mature community tools while keeping the global-fit logic centralized.

## 3. Inference workflow and statistical machinery

A typical GAMBIT workflow proceeds by defining a BSM model in the model database, computing predictions through the relevant modules, calling external backends for specialized calculations, constructing likelihoods from theoretical predictions and experimental measurements, scanning parameter space with ScannerBit, and then producing posterior or profile-likelihood results [2105.03165]. In the scalar singlet dark matter example, the total likelihood is described in the standard composite form
\[
\mathcal{L}_{\text{total}} = \prod_i \mathcal{L}_i ,
\]
with individual contributions from relic density, direct detection, gamma rays, solar neutrinos, Higgs invisible width, and nuisance parameters [1611.05065].

The corpus supports both frequentist and Bayesian inference. The axion preview explicitly contrasted Bayesian posteriors, which capture prior-volume effects and fine-tuning in the pre-inflationary Peccei–Quinn scenario, with frequentist profile likelihoods, which should remain comparatively insensitive to such prior-volume effects [1710.11138]. The first framework release also described ScannerBit support for multiple scanning algorithms, including MultiNest, GreAT, T-Walk, Diver, and postprocessing workflows [1705.07908].

Profile likelihood is the dominant statistical language in many flagship applications. The SUSY conference summaries repeatedly describe profile-likelihood surfaces with \(1\sigma\), \(2\sigma\), \(68\%\), \(95\%\), and sometimes \(99.7\%\) contours, while the scalar singlet analysis reports profiled composite likelihood in the \((M_S,\lambda_{hS})\) plane after profiling over nuisance parameters [1710.02503]. This operational emphasis on likelihood surfaces, rather than single-channel exclusions, is one of the defining features of the corpus.

## 4. Modules, backends, and automation layers

The major physics modules include **ColliderBit**, **DarkBit**, **FlavBit**, **SpecBit**, **DecayBit**, **PrecisionBit**, **NeutrinoBit**, **CosmoBit**, and **ScannerBit** [2105.03165]. ColliderBit implements collider observables and likelihoods from LHC and LEP searches, and the original module description emphasizes a fast recasting chain consisting of cross-section calculation, Monte Carlo event generation, and detector simulation. For large scans, SUSY cross sections are computed with **Pythia8**, event generation is parallelized, and detector simulation is performed by **BuckFast**, a custom fast-simulation tool based on four-vector smearing [1611.05065].

DarkBit supplies relic-density calculations, direct-detection likelihoods, gamma-ray likelihoods, and solar-neutrino likelihoods. Its process catalog stores annihilation and decay information for BSM particles, and if an annihilation produces another BSM particle, DarkBit can simulate its decay “on the fly” with a **Fast Cascade Monte Carlo** [1611.05065]. The dedicated DarkBit overview describes its integration with **DDCalc** for direct detection, **gamLike** for gamma rays, and **nulike** for solar-neutrino likelihoods, while also emphasizing the rescaling factor
\[
f = \frac{\Omega_{\tilde \chi^0_1}}{\Omega_c}
\]
for models in which the neutralino is only a subcomponent of the total dark matter [1711.00463].

CosmoBit extends the corpus into cosmology by providing cosmological observables and likelihoods through interfaces to **AlterBBN**, **CLASS**, **DarkAges**, **MontePython**, **MultiModeCode**, and **plc**, making it possible to combine Big Bang nucleosynthesis, cosmic microwave background, large-scale structure, supernova, and related constraints with particle-physics likelihoods in the same scan [2009.03286]. This substantially enlarges the scope of the corpus beyond collider and dark-matter phenomenology.

A major automation layer is **GUM**, the **GAMBIT Universal Model Machine**, introduced to auto-generate GAMBIT code for new BSM theories from **FeynRules** and **SARAH** inputs [2105.03165]. The full GUM paper describes it as the automation layer from Lagrangians to likelihoods: it reads a `.gum` YAML file and symbolic model definitions, extracts particles and parameters, updates GAMBIT Core, writes module code for SpecBit, DecayBit, DarkBit, and ColliderBit, generates frontend wrappers for backends such as **CalcHEP**, **micrOMEGAs**, **SPheno**, **Pythia**, and **Vevacious**, and emits build instructions together with a sample scan YAML [2107.00030]. This makes the corpus extensible at the level of model onboarding rather than only at the level of analysis reuse.

## 5. Flagship applications in supersymmetry and dark matter

The earliest flagship applications used GAMBIT for frequentist global fits of the **CMSSM**, **NUHM1**, **NUHM2**, and weak-scale **MSSM7**. These analyses extended earlier studies in the number of observables included, the scanning techniques employed, and the treatment of nuisance parameters, and they found that stop co-annihilation provides the best fit in the GUT-scale models, while the MSSM7 best-fit scenario has light higgsinos and highly under-abundant relic density due to efficient chargino co-annihilation [1710.02503]. A subsequent summary emphasized that the viable likelihood landscape is structured by several distinct mechanisms—Higgsino dark matter with chargino co-annihilation, stop co-annihilation, heavy Higgs resonance, and, in some models, stau or sbottom co-annihilation—rather than by a single isolated optimum [1805.06049].

A particularly important collider application is the electroweakino-sector study of the MSSM performed in the four-parameter **EWMSSM**, with free parameters
\[
M_1,\quad M_2,\quad \mu,\quad \tan\beta .
\]
In that setup, all superpartners except the neutralinos and charginos are taken to be heavy enough to decouple from present collider searches [1905.10300]. The analysis used **SpecBit**, **DecayBit**, **ColliderBit**, and **ScannerBit**, scanned the EWMSSM with the **Diver** differential-evolution sampler, and at each sampled point performed full Monte Carlo simulation of the relevant SUSY searches before constructing a joint likelihood from predicted signal rates [1905.10300]. After the initial scan, the collaboration reran the LHC simulations with much higher event statistics, reaching up to **64 million events per point** for the **500 highest-likelihood points**, in order to sharpen the profile likelihood and confidence regions [1905.10300].

The electroweakino fit incorporated recent **13 TeV ATLAS and CMS searches for charginos and neutralinos with 36 fb\(^{-1}\)**, together with invisible decay widths of the \(Z\) and Higgs bosons and LEP SUSY cross-section limits [1905.10300]. Its central conclusion was that current results do **not** impose a generic exclusion of light neutralino and chargino masses in the full EWMSSM interpretation. Instead, the combined data prefer a specific low-mass pattern, with a lightest neutralino below about **200 GeV** at the \(2\sigma\) level and heavier neutralinos below about **700 GeV** in the favored region [1905.10300]. Using the best-fit EWMSSM point as the signal hypothesis, the authors quoted a **local significance of \(3.3\sigma\)** for the combined 13 TeV excesses, reduced to **\(2.9\sigma\)** when selected 8 TeV analyses were added in post-processing [1905.10300].

The scalar singlet dark matter study served a different role: it was both a physics result and a proof of concept for the composite-likelihood machinery. That fit explored a **15-dimensional parameter space**, varying \(M_S\) from **45 to 10,000 GeV**, \(\lambda_{hS}\) from **0.0001 to 10**, and profiling over nuisance parameters [1611.05065]. The preliminary result was that the model was strongly constrained and survived mainly in a narrow Higgs-pole resonance strip near \(M_S \sim 63~\mathrm{GeV}\) [1611.05065].

## 6. Extension beyond supersymmetry

The GAMBIT corpus was designed to be theory-agnostic, and later applications broadened its domain well beyond SUSY. The axion program used GAMBIT for global fits of QCD axions and axion-like particles across laboratory searches, astrophysical bounds, and cosmological constraints. That implementation covered generalized QCD axion models, ALP models, and more specific KSVZ and DFSZ realizations, while also emphasizing the distinction between Bayesian and frequentist inference in the presence of the initial misalignment angle \(\theta_\mathrm{i}\) [1710.11138].

Extended Higgs sectors form another strand of the corpus. The preliminary global fit of the CP-conserving type-II 2HDM combined theoretical constraints, Higgs searches at colliders, electroweak precision observables, and flavour constraints within a frequentist GAMBIT analysis using **ScannerBit** and **Diver** [2007.11942]. A notable result of that study was that the measurements of \(B\rightarrow K^*\mu^+\mu^-\) angular observables cannot be explained in the type-II 2HDM [2007.11942].

Cosmology was incorporated systematically through CosmoBit, which brought flat \(\Lambda\)CDM, inflationary models, neutrino-mass scenarios, non-standard radiation content, exotic energy injection from annihilating or decaying dark matter, and neutron-lifetime variations into the same global-fit environment [2009.03286]. The first application showed that if one allows both a modified neutrino temperature \(r_\nu\) and extra ultra-relativistic species \(\Delta N_{\rm ur}\), then the usual cosmological upper bound on \(\sum m_\nu\) becomes much weaker than in the standard scenario [2009.03286]. This illustrates the broader corpus logic: cosmological and particle-physics assumptions can be varied jointly rather than imposed sequentially.

## 7. Methodological significance and recurrent themes

Several general lessons recur across the corpus. One is that **simplified-model exclusions are not directly transferable** to more realistic theories. The electroweakino study states three reasons: heavier neutralinos and charginos introduce additional signal processes that can improve the fit elsewhere, heavier states can change which signal region is deemed most sensitive, and bino/wino/higgsino mixing alters production cross sections and branching ratios even at the same light-state masses [1905.10300]. This is a methodological warning against reading a two-mass exclusion contour as a generic statement about the full MSSM electroweakino sector.

A second recurrent theme is the treatment of relic density as an **upper limit** rather than a strict equality. The SUSY global fits repeatedly defined the neutralino fraction through
\[
f = \frac{\Omega_\chi}{\Omega_{\rm DM}},
\]
and rescaled direct-detection quantities accordingly, allowing the neutralino to be only a subcomponent of the observed dark matter [1805.06049]. The first SUSY results emphasized that this conservative choice materially affects the interpretation of co-annihilation and funnel regions, while the scalar singlet example likewise rescaled the dark-matter density used in dwarf and local-Galaxy calculations according to the predicted relic density at each point [1710.02503].

A third theme is computational scale. The electroweakino analysis noted that adding selected 8 TeV searches to the initial scan would have roughly doubled the computational cost, which is why they were included only in post-processing [1905.10300]. The same study’s high-statistics reruns, up to 64 million events per point, exemplify the computational burden that global collider fits can entail. This suggests that the corpus is not merely a library of likelihood functions; it is also an infrastructure for managing expensive end-to-end simulations.

Taken together, these features explain why the 2021 overview characterized the GAMBIT corpus as a reusable computational corpus for modern particle-physics model testing, and why it described GAMBIT itself as not just one code package but an evolving ecosystem [2105.03165]. Its enduring significance lies in providing a common architecture in which theory calculation, decay and production prediction, Monte Carlo event generation, backend interoperability, and statistical inference can be assembled into a single global-fit framework.

Source: https://www.emergentmind.com/topics/gambit-corpus