Excalibur: Diverse Scientific Systems

• Excalibur is a collective label for multiple independent research systems, each with distinct architectures and scientific objectives across varied domains.
• Each system employs tailored methodologies—from Compton polarimetry and PCA-based wavelength calibration to interactive SVM learning and vectorization tuning—to address domain-specific challenges.
• Practical insights reveal that targeted technical refinements, such as controlled systematics in polarimetry and effective task difficulty assessment in cybersecurity, drive significant performance improvements.

Excalibur is a recurring name in contemporary research, but in the cited literature it does not denote a single instrument, algorithm, or programme. Instead, it labels several technically distinct systems: X-Calibur, a hard X-ray polarimeter; ExCALIBUR, the UK exascale programme that funded a RISC-V HPC testbed; EXCALIBUR, a catalogue-oriented exoplanet transit-spectroscopy pipeline; Excalibur, a non-parametric wavelength-calibration framework for precision spectrographs; Excalibur, the “Experiment for Calibration with Uranium” neutron source for warhead-verification research; Excalibur, a forensic cross-modal image-retrieval system; and Excalibur / PentestGPT v2, a difficulty-aware penetration-testing agent (Beilicke et al., 2011, Abarr et al., 2022, Brown et al., 2023, Brown, 2024, Mugnai et al., 2024, Zhao et al., 2020, Jeon et al., 17 Mar 2025, Böhne et al., 2023, Deng et al., 19 Feb 2026).

1. Nomenclature and domain scope

The capitalization and hyphenation are technically significant. In the supplied literature, the name appears in astrophysical instrumentation, HPC infrastructure, exoplanet data reduction, precision spectrograph calibration, nuclear-verification experimentation, forensic image retrieval, and autonomous cybersecurity.

Variant	Domain	Role
X-Calibur	Hard X-ray astrophysics	Compton polarimeter
ExCALIBUR	HPC / exascale	UK programme funding RISC-V testbed
EXCALIBUR	Exoplanet spectroscopy	Automated catalogue pipeline
Excalibur	Precision spectroscopy	Wavelength-calibration framework
Excalibur	Warhead verification	Reconfigurable neutron source
Excalibur	Digital forensics	Cross-modal image retrieval system
Excalibur / PentestGPT v2	Cybersecurity	Difficulty-aware pentesting agent

A common misconception is to treat these as variants of one project. The cited work supports the opposite conclusion: the shared label masks a set of independent research efforts with distinct architectures, metrics, and scientific or operational objectives.

2. X-Calibur in hard X-ray polarimetry

X-Calibur was introduced as a hard X-ray polarimeter for measuring the linear polarization of astrophysical X-rays in the 10–80 keV band in the focal plane of the InFOCuS / InFOCμS grazing-incidence hard X-ray telescope. Its measurement concept is Compton polarimetry: linearly polarized photons preferentially scatter perpendicular to the electric-field vector, so the instrument reconstructs the azimuthal scattering angle from a low-Z central scatterer and surrounding high-Z CZT absorbers. In the 2011 design paper, the central element is an EJ-200 scintillator rod read out by a Hamamatsu R7600U-200 PMT; the absorber assembly consists of $2 \times 2\,\mathrm{cm^2}$ CZT modules segmented into 64 pixels with 2.5 mm pitch, using 0.2 cm or 0.5 cm detector thicknesses. The design emphasized high detection efficiency close to unity, low background, and controlled systematics through coincidence triggering, shielding, and rotation. Simulations for a Crab-like source gave $\mu \approx 0.52$ for a 100% polarized beam and a projected MDP about 4% in the 10–80 keV band for 5.6 hour on-source exposure plus 1.4 hours of background. Laboratory tests with $^{152}\mathrm{Eu}$ calibration lines and a polarized 288 keV beam generated from $^{137}\mathrm{Cs}$ scattering measured the expected 0.22 relative modulation amplitude; the unpolarized acceptance measurement showed a 90° modulation attributable to the fourfold detector geometry rather than to source polarization (Beilicke et al., 2011).

By the 2018/19 Antarctic long-duration balloon flight, X-Calibur had matured into a balloon-borne telescope operating in the $\sim 15$–$50\;\mathrm{keV}$ band. The flight instrument used the InFOC$\mu$S mirror with $8\;\mathrm{m}$ focal length, $40\;\mathrm{cm}$ diameter, and effective area of about $93\;\mathrm{cm^2}$ at $\mu \approx 0.52$0 and $\mu \approx 0.52$1 at $\mu \approx 0.52$2. The focal-plane polarimeter employed an $\mu \approx 0.52$3 long, $\mu \approx 0.52$4 diameter beryllium scattering rod, sixteen side CZT modules plus a seventeenth forward module, and 1088 channels, with the full assembly rotating at 3 rpm. The mean modulation factor in the band of interest was $\mu \approx 0.52$5, the on-source pointing stability was 1.0–3.6 arcsec (3$\mu \approx 0.52$6), and the single-pixel background rate in the 15–50 keV polarization band was about 2.1 Hz after shield-veto subtraction. The mission was reported as nominal overall, although the anticoincidence threshold was higher than expected and the readout chain exhibited unanticipated dead time. Even with a flight duration of less than three days, the instrument obtained the first hard X-ray polarization constraints for the accretion-powered pulsar GX 301–2, and the measured performance directly informed the follow-up XL-Calibur design (Abarr et al., 2022).

3. ExCALIBUR as a UK exascale HPC programme context

In the HPC literature, ExCALIBUR refers to the UK ExCALIBUR HES / H&ES exascale programme, specifically as the funding mechanism behind a free-access RISC-V HPC testbed stood up in early 2022. The testbed was designed to feel like a familiar HPC service rather than a bare-board environment: it provided a login node, a module environment, Slurm, and a shared filesystem, and it supported scientific workloads such as MONC, WRF, and CP2K over a period of 12 months. Hardware diversity was a central feature, with six types of physical RISC-V CPUs plus soft-core RISC-V CPUs deployed via FPGAs. The reported experience was that much of the “basic HPC plumbing” worked “pretty much out of the box,” including Linux boot, NFS, and Slurm. The software ecosystem already included PETSc, FFTW, HDF5, NetCDF, MPICH, and OpenMPI, so MPI codes could run across RISC-V nodes in the same way as on other architectures. The largest challenge, however, was software tooling, especially vectorization: mainline GCC did not support RISC-V vectorization because the v0.7.1 branch had been dropped, T-Head’s GCC supported v0.7.1, and Clang supported only v1.0. To bridge that mismatch, the team built a tool that rewrote Clang-generated v1.0 vector assembly back to v0.7.1. In Polybench experiments compiled with Clang 16, vectorization was sufficiently important that for Heat-3D a single vectorized core on the $30 Allwinner D1 outperformed the other CPUs with 2 or 4 cores (Brown et al., 2023).

The later extended abstract generalizes these experiences into an assessment of RISC-V for HPC. It records a hardware progression from HiFive Unmatched, MangoPi MQ-Pro, VisionFive V1, and VisionFive V2 to the Sophon SG2042, which offers 64 cores and up to 128GB memory on the Milk-V Pioneer workstation. The authors describe SG2042 as “for the first time a serious RISC-V option for HPC workloads,” while also noting that it remains around between four and eight times slower than modern x86-based hardware commonly used in HPC. The technical gaps identified as most urgent are better vectorization support, better performance profiling tools, improved access to hardware performance counters, and a clearer HPC value proposition. On profiling specifically, the platform lacks a high-level tool comparable to CrayPat, Vtune, or Forge. The paper also highlights possible future directions through RVV v1.0 hardware such as Sophon SG2044 and through accelerators including Esperanto ET-SoC, InspireSemi Thunderbird, and Tenstorrent Greyskull and Wormhole (Brown, 2024).

4. EXCALIBUR in astronomical data reduction and calibration

In exoplanetary science, EXCALIBUR is a transit-spectroscopy reduction pipeline associated with the Roudier et al. (2021) catalogue of HST/WFC3 transmission spectra. Its distinguishing characteristics are fully automated, uniform processing and the use of persistent intermediate data products, which preserve the chain of inference and make the processing more traceable. In the catalogue-level comparison study, EXCALIBUR is evaluated against Iraclis and CASCADe using HST/WFC3 spectra binned to $\mu \approx 0.52$7 at $\mu \approx 0.52$8 with 25 spectral points per spectrum. The EXCALIBUR catalogue contains 62 spectra, with 30 planets overlapping Tsiaras et al. (2018) / Iraclis and 6 planets overlapping Edwards et al. (2022) / Iraclis; the study ultimately compares 35 planets, including 22 shared across the three main catalogue comparisons. The principal result is not a simple mean offset but a systematic difference in spectral shapes and retrieval outcomes: mean ratios are centered near 1, yet EXCALIBUR vs CASCADe has 14.29% of ratios outside a ±25% tolerance in one comparison; EXCALIBUR uncertainties are typically about ~15% larger than CASCADe’s; only about 40% of spectral bins are consistent within 1σ and about 68% within 2σ. In uniform retrievals using Alfnoor / TauREx 3, EXCALIBUR often explores higher temperatures, roughly 2000–3000 K, broader CH$\mu \approx 0.52$9 and H$^{152}\mathrm{Eu}$0O abundance ranges, and cloud pressures often around 10,000 Pa, with corresponding changes in ADI and in whether a planet sits in a “water detected” or “no-detection” regime. The paper’s central conclusion is that pipeline-induced bias is not reflected in the quoted uncertainties (Mugnai et al., 2024).

A second astronomical use of the name appears in precision spectroscopy. Here Excalibur is a non-parametric, hierarchical wavelength-calibration framework for extreme-precision radial-velocity work, motivated by the need to calibrate or stabilize spectrographs to better than about $^{152}\mathrm{Eu}$1 pixels. Its core model is a low-dimensional representation of calibration states,

$^{152}\mathrm{Eu}$2

where $^{152}\mathrm{Eu}$3 is the fiducial calibration, $^{152}\mathrm{Eu}$4 are basis functions, and $^{152}\mathrm{Eu}$5 are exposure-dependent amplitudes. The implementation described in the paper uses PCA for dimensionality reduction and PCHIP for order-by-order interpolation, exploiting the dense, stable calibration points delivered by laser frequency combs or etalons. Demonstrated on EXPRES, with 1227 LFC exposures and 78 ThAr exposures over 29 unique nights, the method assigns wavelengths to comb lines with an RMS residual about five times lower than polynomial fits to individual exposures. Applied to radial-velocity measurements of HD 34411, it reduced the RMS scatter over a 10-month time baseline from $^{152}\mathrm{Eu}$6 to $^{152}\mathrm{Eu}$7. The same paper also makes the principal limitation explicit: the interpolation-based approach performs much worse with sparse calibration sources such as ThAr (Zhao et al., 2020).

5. Excalibur as the “Experiment for Calibration with Uranium” neutron source

In warhead-verification research, Excalibur means Experiment for Calibration with Uranium, a reconfigurable neutron source built to support two distinct use cases: neutron radiography / transmission imaging, which requires a directed beam of high-energy neutrons, and fissile-isotope / induced-fission detection, which benefits from a softer sub-MeV neutron spectrum. At its core is a commercial Thermo Fisher P-385 deuterium-tritium neutron generator specified and measured to produce up to $^{152}\mathrm{Eu}$8 neutrons/s with the A3082 tube and up to $^{152}\mathrm{Eu}$9 neutrons/s with the A3083 tube. The generator output scales linearly with beam current and approximately as $^{137}\mathrm{Cs}$0 with accelerating voltage. The generator is housed inside a 32-inch diameter, 23.62-inch high carbon-steel cylinder that moderates the mean neutron energy to under 500 keV; this is surrounded by 5% borated polyethylene, giving an overall 48" × 48" assembly that is 30" tall. The generator is inclined 10° backward to avoid a steep flux gradient, and source monitoring is provided by an Eljen-410 ZnS fast-neutron detector.

The two operating modes define the system’s technical identity. In collimated mode, a 17.5° tapered aperture with 2-inch height allows direct streaming of 14 MeV neutrons for radiography. At 1.4 m from the source, the 10–20 MeV flux is nearly flat from beam axis to 5° off-axis, falls to somewhat less than half at 8.75°, and drops by more than 95% by 12°; neutrons in the 10–20 MeV range account for 70% of the total flux, and the overall flux is 1.6 times greater than in moderated mode. In moderated mode, the DT generator is fully enclosed by steel while most of one polyethylene face is removed, producing a broad sub-MeV field for fissile-isotope detection. At 2 m from the source, the low-energy flux varies smoothly with angle and is still about half its on-axis value at slightly over 25° off-axis; neutrons below 0.8 MeV make up 89% of the total flux, the flux below 0.8 MeV is 4.5 times higher than in collimated mode for the same source rate, and the 0–0.8 MeV to 10–20 MeV flux ratio is 274 times higher than in collimated mode. Measurements with the BTI N-Probe, the Nested Neutron Spectrometer, and a 2-inch NaI(Tl) detector agreed well with MCNP6 modeling in both angular and spectral behavior, including gamma-ray features at 57 keV and 203 keV and simulated contributions from $^{137}\mathrm{Cs}$1B and $^{137}\mathrm{Cs}$2Fe. Reconfiguration between the two source states can be completed in under one hour without special tools or lifting equipment (Jeon et al., 17 Mar 2025).

6. Excalibur in retrieval systems and autonomous security

In digital forensics, Excalibur is a zero-shot cross-modal image-retrieval system extended with interactive learning. It is designed for large, unstructured image collections in which investigators may begin from either natural-language text or an example image. The system uses CLIP to embed both queries and images into a shared space and ranks candidates by cosine similarity,

$^{137}\mathrm{Cs}$3

Interactive refinement is implemented with an SVM using scikit-learn’s SVC: the user marks retrieved images as relevant or non-relevant, the classifier is trained on the corresponding embeddings, and the corpus is re-ranked by classifier confidence. The simulation study used 20 forensic-relevant scene categories from Places365, each with 100 validation images, and evaluated performance by $^{137}\mathrm{Cs}$4 and $^{137}\mathrm{Cs}$5. Natural-language queries outperformed image queries initially, with MAP@50 of 21.4 versus 15.1 and Recall@200 of 28.3 versus 18.5. After 10 interaction rounds, performance improved to 67.5 and 50.5 for natural-language queries and to 61.4 and 45.9 for image queries. The authors report that the zero-shot method with interactive learning approaches a supervised ResNet50 baseline in Recall@200 and outperforms it in MAP@50. A user study with five investigators working on imagery from concluded investigations found that participants preferred image queries for specific domain semantics such as container seal, but preferred natural-language queries for broad exploration such as sea port before using feedback to “zoom in.” The participants rated the system 9.4/10 on the System Usability Scale and expressed interest in daily use (Böhne et al., 2023).

In cybersecurity, Excalibur / PentestGPT v2 denotes a difficulty-aware LLM penetration-testing agent. The paper first analyzes 28 LLM-based pentesting systems and then evaluates five representative implementations across XBOW, the PentestGPT Benchmark, and GOAD. Its central distinction is between Type A failures, caused by missing tools, prompts, parsing, or documentation, and Type B failures, caused by poor planning, state management, premature commitment, and inability to estimate task difficulty. Excalibur addresses the first class with a Tool and Skill Layer comprising typed interfaces for 38 tools across six categories and RAG over tool documentation, CVE descriptions, and attack playbooks. It addresses the second with Task Difficulty Assessment (TDA) and Evidence-Guided Attack Tree Search (EGATS). TDA combines four measurable dimensions—horizon estimation, evidence confidence, context load, and historical success—into a Task Difficulty Index,

$^{137}\mathrm{Cs}$6

with $^{137}\mathrm{Cs}$7 and $^{137}\mathrm{Cs}$8. EGATS then uses these scores to choose between reconnaissance and exploitation, prioritize branches, backtrack, and prune. The reported results are up to 91% task completion on XBOW, 39–49% relative improvement over baselines, 12 of 13 machines rooted on the PentestGPT Benchmark in some settings, and 4 of 5 hosts compromised on GOAD versus 2 for prior systems. In ablation, the progression from Base to + Tool Layer to + TDA-EGATS to + Memory improves scores from 54 / 8 / 2 to 68 / 9 / 2, 77 / 11 / 3, and 85 / 12 / 4 on XBOW / Pentest-Ben / GOAD, respectively. The paper’s main claim is that model scaling alone is insufficient: durable gains arise from explicit difficulty-aware planning rather than from stronger base models alone (Deng et al., 19 Feb 2026).