VERITAS: Research & Applications Overview

• VERITAS is a multi-domain designation encompassing a high-energy gamma-ray observatory, binary vulnerability detection, clinical hypothesis testing, deepfake analysis, and theorem proving.
• The astronomical VERITAS observatory features four Davies–Cotton telescopes with advanced trigger systems and performance upgrades that significantly enhance sensitivity and resolution.
• In cybersecurity and AI, VERITAS systems employ verification-driven methodologies such as static slicing, Monte Carlo tree search, and evidence fusion to validate results and reduce errors.

VERITAS denotes several distinct research entities in contemporary scholarship. The dominant historical usage is the Very Energetic Radiation Imaging Telescope Array System, a ground-based $\gamma$-ray observatory at the Fred Lawrence Whipple Observatory in southern Arizona that operates from $\sim 85$ GeV to $>30$ TeV and has been in science operation since 2007 (Park, 2015, Hanna et al., 13 Jul 2025). In more recent arXiv literature, the same name or closely related capitalization variants also identify verification-oriented systems for binary vulnerability detection, multimodal clinical hypothesis testing, multimodal data refinement, synthetic-image and deepfake analysis, automated fact-checking, and zero-shot formal theorem proving (Zheng et al., 14 May 2026, Stoffl et al., 13 Apr 2026, Xu et al., 17 Oct 2025, Srivastava et al., 7 Jul 2025, Tan et al., 28 Aug 2025, Rothermel et al., 13 Jan 2026, Acharya et al., 17 Jun 2026).

1. Research uses of the designation

Usage	Domain	Core function
Very Energetic Radiation Imaging Telescope Array System	VHE $\gamma$-ray astronomy	Four-telescope IACT observatory
Veritas	Binary security	Memory-corruption vulnerability detection in stripped binaries
VERITAS	Clinical AI	Image-derived hypothesis testing with auditable evidence
VERITAS	Multimodal data curation	Vision-prior-guided SFT-data refinement
VERITAS	Synthetic-image forensics	Real/fake detection with artifact localization and explanation
Veritas	Deepfake detection	Pattern-aware MLLM reasoning on HydraFake
VeriTaS	Fact-checking benchmark	Dynamic multimodal AFC benchmark
VERITAS	Formal theorem proving	Verifier-guided proof search

The astronomy usage is documented across a long sequence of VERITAS observatory papers from 2009 through 2025, whereas the AI- and systems-oriented usages appear in 2025–2026 preprints and are generally centered on explicit grounding, executable evidence, or verifier feedback (Weekes et al., 2010, Park, 2015, Hanna et al., 13 Jul 2025, Zheng et al., 14 May 2026, Stoffl et al., 13 Apr 2026, Rothermel et al., 13 Jan 2026, Acharya et al., 17 Jun 2026).

2. The astrophysical VERITAS observatory: design and evolution

The astronomical VERITAS is a stereoscopic imaging atmospheric Cherenkov telescope array at the Fred Lawrence Whipple Observatory in southern Arizona. In the 2025 observatory chapter, it is described as four 12 m imaging atmospheric-Cherenkov telescopes situated at $31^\circ 40^\prime$ N, $110^\circ 57^\prime$ W, $1270$ m asl, following a classic Davies–Cotton optical design (Hanna et al., 13 Jul 2025). Each reflector is a tessellated dish composed of 345 hexagonal, front-surface aluminum-coated glass facets with 24 m focal length, and each telescope camera contains 499 close-packed 26 mm photomultiplier tubes equipped with Winston cones, giving a $3.5^\circ$ field of view and an angular pixel size of $0.15^\circ$ (Hanna et al., 13 Jul 2025). Earlier overview papers describe comparable array-level parameters: four 12 m Davies–Cotton telescopes, mirror area $\simeq 110\ \mathrm{m}^2$ per dish, and typical inter-telescope baselines of $\sim 85$0–$\sim 85$1 m (Staszak et al., 2015, Benbow, 2019).

The trigger and readout chain is explicitly hierarchical. Level 1 uses constant-fraction discriminators on each PMT channel; level 2 applies nearest-neighbor pattern logic such as $\sim 85$2 adjacent PMTs within $\sim 85$3 ns; level 3 requires an array coincidence of $\sim 85$4 telescopes within $\sim 85$5 ns, with real-time compensation of geometrical delays (Hanna et al., 13 Jul 2025). Upon an L3 trigger, 32 ns windows of 2 ns-digitized waveforms are read out from each telescope’s flash ADCs into a central event builder, with typical trigger rates of $\sim 85$6 Hz and $\sim 85$7 deadtime at 400 $\sim 85$8s readout per event (Hanna et al., 13 Jul 2025).

The instrument’s hardware history is usually partitioned into three epochs. The performance paper identifies V4 (2007–08/2009) as the original array, V5 (09/2009–08/2012) as the configuration after relocation of telescope 1 to a more symmetric position, and V6 (09/2012–present) as the configuration after the high–quantum-efficiency PMT camera upgrade (Park, 2015). The 2009 relocation improved stereo baselines and yielded a $\sim 85$9 gain in sensitivity below $>30$0 TeV; the 2012 upgrade increased photon collection by $>30$1 (Park, 2015). Other status papers also emphasize the 2011–2012 trigger and PMT upgrades, including replacement of the original PMTs with high-quantum-efficiency models having peak photon detection efficiency of $>30$2 or quantum efficiency of $>30$3–$>30$4 (Staszak, 2013, Galante, 2012).

3. Instrument performance, calibration, and event analysis

The observatory’s operating range is consistently given as $>30$5 GeV to $>30$6 TeV or $>30$7 GeV to $>30$8 TeV, with an analysis threshold of $>30$9 GeV under standard low-zenith dark-sky conditions (Park, 2015, Hanna et al., 13 Jul 2025, Benbow, 2019). The canonical point-source sensitivity is the ability to detect a source with flux $\gamma$0 of the Crab Nebula flux within 25 hours, or equivalently at $\gamma$1 in $\gamma$2 h (Park, 2015, Hanna et al., 13 Jul 2025, Staszak et al., 2015). Differential sensitivity is defined as the minimum flux detectable at $\gamma$3 in 50 h, subject to at least 10 $\gamma$4-ray excess events in each energy bin (Park, 2015, Hanna et al., 13 Jul 2025).

The standard significance estimator is the Li–Ma likelihood-ratio expression. In the observatory performance paper it is written as

$\gamma$5

with $\gamma$6 the ON/OFF normalization, typically $\gamma$7 in wobble mode (Park, 2015). VERITAS analyses use Hillas-moment parametrization, stereoscopic reconstruction from multiple telescope images, and $\gamma$8–hadron separation through scaled width/length cuts, timing parameters, and, in some later analyses, boosted decision trees, template likelihood fits, or two-dimensional Gaussian image fitting (Staszak et al., 2015, Patel, 2022).

Representative performance figures for V6 with medium cuts are explicit. The differential sensitivity is $\gamma$9 Crab at 200 GeV, $31^\circ 40^\prime$0 Crab at 1 TeV, and $31^\circ 40^\prime$1 Crab above 10 TeV (Park, 2015). The effective collection area is $31^\circ 40^\prime$2 at 100 GeV, $31^\circ 40^\prime$3 at 1 TeV, and $31^\circ 40^\prime$4 above 10 TeV (Park, 2015). The 68% containment radius is $31^\circ 40^\prime$5 at 200 GeV, $31^\circ 40^\prime$6 at 1 TeV, and $31^\circ 40^\prime$7 above a few TeV (Park, 2015). The energy resolution is $31^\circ 40^\prime$8 at 200 GeV, improves to $31^\circ 40^\prime$9 near 1 TeV, and degrades to $110^\circ 57^\prime$0 at $110^\circ 57^\prime$1 TeV (Park, 2015).

Calibration is likewise explicit. Nightly LED flasher runs establish PMT gains and timing offsets, pedestal and baseline corrections are applied, and optical throughput is tracked through mirror-reflectivity measurements and periodic muon-ring studies (Hanna et al., 13 Jul 2025, Patel, 2022). The residual cosmic-ray background rate falls from $110^\circ 57^\prime$2 Hz before cuts to $110^\circ 57^\prime$3 Hz after standard box cuts, and systematic uncertainties on flux are $110^\circ 57^\prime$4, dominated by atmosphere, mirror reflectivity, and PMT calibration (Park, 2015).

The major upgrades produced quantitative gains. At 200 GeV, the 50 h sensitivity improved from $110^\circ 57^\prime$5 Crab in V4 to $110^\circ 57^\prime$6 Crab in V5 and $110^\circ 57^\prime$7 Crab in V6; the $110^\circ 57^\prime$8-Crab detection time improved from 35 h to 29 h to 25 h (Park, 2015). This is the core instrumental basis for the observatory’s long science program.

4. Scientific program and legacy of the observatory

VERITAS’s science portfolio spans extragalactic sources, Galactic accelerators, dark matter, cosmic rays, and multimessenger follow-up. The 2025 chapter states that the observatory has accumulated over sixteen years of observations and has produced key constraints on cosmic-ray origins, dark matter, intergalactic radiation fields, and fundamental physics (Hanna et al., 13 Jul 2025). It further states that VERITAS has detected $110^\circ 57^\prime$9 blazars up to $1270$0, measured the Crab Nebula spectrum to 40 TeV, and constrained the intergalactic magnetic field to $1270$1–$1270$2 G on Mpc scales through non-detection of magnetically broadened pair cascades (Hanna et al., 13 Jul 2025).

The AGN program is especially prominent. A 2019 summary reports $1270$3 hours of observations targeted on active galactic nuclei, approximately 300 AGN observed, and 39 detections, with most detections accompanied by contemporaneous broadband observations (Benbow, 2019). A 2021 update reports nearly 7,000 hours on AGN, approximately 300 AGN observed, and 40 detections (Benbow, 2021). Specific results include the 5.8$1270$4 detection of TXS 0506+056 after the IceCube neutrino alert IC-170922A, Ton 599 at 10$1270$5, and the 9$1270$6 detection of 3C 264 (Benbow, 2019). Deep AGN monitoring has also yielded minute-scale variability and multiwavelength constraints on emission-region sizes and Doppler factors (Benbow, 2021, Hanna et al., 13 Jul 2025).

Galactic results include supernova remnants, pulsars, pulsar wind nebulae, binaries, and Galactic-center observations. The 2025 chapter states that VERITAS achieved the first IACT detection of pulsed emission from the Crab pulsar up to 400 GeV, indicating emission zones near or beyond the light cylinder and disfavoring pure curvature radiation (Hanna et al., 13 Jul 2025). It also identifies MGRO J1908+06 and VER J2227+608 (G106.3+2.7) as promising Galactic PeVatron candidates because their spectra show no cutoff up to $1270$7 TeV (Hanna et al., 13 Jul 2025). A 2022 review highlights Sgr A* detected at 38$1270$8 in 125 h of large-zenith-angle observations, Cas A with a joint VERITAS+Fermi-LAT cutoff of $1270$9 TeV, and the 7.8$3.5^\circ$0 detection of 3C 264 above 315 GeV (Patel, 2022).

Indirect dark-matter searches are another long-running program. The 2025 chapter reports that a combined 230 h exposure on four dwarf spheroidal galaxies yields 95% C.L. upper limits on the velocity-averaged WIMP annihilation cross section down to $3.5^\circ$1 at 1 TeV (Hanna et al., 13 Jul 2025). Earlier science highlights describe a joint dSph analysis with $3.5^\circ$2 h on five dwarf spheroidals and limits of $3.5^\circ$3 and $3.5^\circ$4 for $3.5^\circ$5 TeV (Staszak et al., 2015).

The observatory has also measured the cosmic-ray electron spectrum from 300 GeV to 5 TeV, finding a spectral break at $3.5^\circ$6 TeV, and has carried out primordial-black-hole searches, multimessenger target-of-opportunity campaigns, and long-baseline monitoring of transients and binaries (Staszak et al., 2015, Hanna et al., 13 Jul 2025). Its reported legacy includes 150+ peer-reviewed publications, a public high-level data catalog called VTSCat, and training of over 70 Ph.D. students (Hanna et al., 13 Jul 2025).

5. Binary vulnerability detection and formal theorem proving

In binary security, "Veritas" is the name of a semantically grounded framework for memory corruption vulnerability detection in stripped binaries (Zheng et al., 14 May 2026). The system is organized as a three-stage pipeline—Static Slicer $3.5^\circ$7 Dual-View LLM Detector $3.5^\circ$8 Multi-Agent Validator—designed to recover interprocedural value flow and object-boundary relations, reason about feasible control-flow and bounds constraints, and then confirm candidates through concrete execution (Zheng et al., 14 May 2026). The slicer operates over RetDec-lifted LLVM IR, reconstructing def-use, calls, returns, globals, and pointer operations into compact witness-backed flow objects; the detector reasons jointly over decompiled C and selective LLVM IR; and the validator uses radare2, Valgrind, breakpoints, and debugger-visible artifacts to confirm or reject candidate bugs (Zheng et al., 14 May 2026). On a dataset of 20 real-world out-of-bounds vulnerabilities from 10 projects, Veritas found 18 of 20 vulnerabilities, yielding 90% recall (Zheng et al., 14 May 2026). In the exhaustive false-positive subset, the validator checked 623 detector candidates, confirmed 9 real bugs, and produced 0 false positives; a real-world application to Apple’s USD library uncovered a previously unknown out-of-bounds read that was confirmed and assigned a CVE (Zheng et al., 14 May 2026).

A different 2026 system uses the same name for verifier-guided proof search for zero-shot formal theorem proving (Acharya et al., 17 Jun 2026). This VERITAS uses a two-phase protocol: Best-of-$3.5^\circ$9 sampling first, then a critic-guided Monte Carlo tree search pass that feeds structured verifier errors back into exploration (Acharya et al., 17 Jun 2026). The framework preserves every theorem solved in Phase 1, and the paper states the monotonicity property $0.15^\circ$0 (Acharya et al., 17 Jun 2026). Rather than reducing the verifier to a binary pass/fail bit, the system retains four signals—syntax valid, type-correct, goal progressed, and proof completed—and injects failures as explicit negative examples into later search (Acharya et al., 17 Jun 2026). Quantitatively, it reaches 40.6% on miniF2F, versus 36.9% for an independently run Best-of-5 and 26.2% for Portfolio, and 7.3% on VERITAS-CombiBench, versus 1.8% for Best-of-5 and 3.6% for Portfolio (Acharya et al., 17 Jun 2026).

These two systems use "verification" in literal operational senses: one through debugger-visible runtime evidence in compiled binaries, the other through structured Lean feedback during proof search.

6. Clinical hypothesis testing with auditable evidence trails

"VERITAS" also names Verifiable Epistemic Reasoning for Image-Derived Hypothesis Testing via Agentic Systems, a multi-agent clinical-research system for testing natural-language hypotheses on multimodal datasets while producing a fully auditable evidence trail (Stoffl et al., 13 Apr 2026). Its workflow is divided into four sequential phases: Analysis Planning, Segmentation, Statistical Analysis, and Interpretation and Verdict (Stoffl et al., 13 Apr 2026). Planning outputs a structured JSON plan specifying testability, target cohorts, structures to segment, derived measurements, the statistical test, and a priori power at a planning SESOI; segmentation uses a promptable backend such as SAT; statistical analysis writes and iteratively debugs Python code that loads masks and metadata via a constrained API, computes derived metrics, performs the prescribed statistical test, and outputs statistical_results.json; interpretation then returns a final verdict $0.15^\circ$1 (Stoffl et al., 13 Apr 2026).

A defining component is the deterministic Evidence Classification Operator. With $0.15^\circ$2 and power $0.15^\circ$3 at a hypothesis-specific SESOI $0.15^\circ$4, VERITAS maps results to four labels: Supported, Refuted, Underpowered, or Invalid (Stoffl et al., 13 Apr 2026). Supported requires $0.15^\circ$5 and matching effect direction; Refuted requires either $0.15^\circ$6 with $0.15^\circ$7 or a significant effect in the opposite direction; Underpowered requires $0.15^\circ$8 and $0.15^\circ$9; Invalid covers untestable hypotheses or execution failures (Stoffl et al., 13 Apr 2026). This explicit separation of non-significance from inadequate power is central to the system’s epistemic design.

The benchmark contains 64 hypotheses across six complexity levels, split evenly across ACDC cardiac MRI with 150 subjects and UCSF-PDGM brain-glioma MRI with 501 subjects (Stoffl et al., 13 Apr 2026). Reported metrics include evidence-label accuracy, verdict accuracy, completion rate, and verifiability rate. In majority vote across 10 runs per hypothesis on testable hypotheses L1–L5, open-weight VERITAS achieves 67.8% evidence accuracy, 71.2% verdict accuracy, 78.1% completion, and 77.3% verifiability; frontier VERITAS achieves 76.3% evidence accuracy, 81.4% verdict accuracy, 87.5% completion, and 86.6% verifiability (Stoffl et al., 13 Apr 2026). The paper emphasizes that even failures remain diagnosable because segmentation masks, code, and JSON outputs are inspectable (Stoffl et al., 13 Apr 2026).

7. Media authenticity, multimodal data quality, and dynamic fact-checking

Several further VERITAS systems operate in multimodal authenticity and data-quality settings. One 2025 pipeline, expanded in its details as "Vision-Priors Evaluation and Refinement through Integration of Tri-Expert Assessment with Shrinkage," is a four-stage SFT-data curation framework for large multimodal models (Xu et al., 17 Oct 2025). It extracts vision priors with RAM++ and PP-OCRv4, obtains critique rationales and scores from GPT-4o, Gemini-2.5-Pro, and Doubao-1.5-pro, fuses scores using domain-aware James–Stein shrinkage, distills the ensemble into a 7 B critic via Group Relative Policy Optimization, and then selects the best rewritten answer from a candidate pool (Xu et al., 17 Oct 2025). Fine-tuning Qwen2-VL on 95,955 refined samples improves six benchmarks, including OCR-VQA from 57.78 to 72.13, MME from 1680.9 to 1695.1, MM-Vet from 50.78 to 57.14, and an average gain of +7.4 points over raw-data SFT (Xu et al., 17 Oct 2025).

A separate 2025 system, "VERITAS: Verification and Explanation of Realness in Images for Transparency in AI Systems," addresses zero-shot detection of AI-generated images from 32×32 inputs (Srivastava et al., 7 Jul 2025). Its five-stage pipeline combines DRCT super-resolution, GradCAM heatmaps, patch weighting with

$\simeq 110\ \mathrm{m}^2$0

CLIP-based artifact scoring with

$\simeq 110\ \mathrm{m}^2$1

and MOLMO-7B-D for human-readable artifact descriptions (Srivastava et al., 7 Jul 2025). The paper emphasizes explanatory rather than purely classificatory evaluation; it cites a ResNet18 baseline on CIFAKE at 92.98% binary real/fake accuracy and argues that VERITAS adds artifact-level filtering and semantic evidence for zero-shot detection of unseen synthetic sources without fine-tuning (Srivastava et al., 7 Jul 2025).

Another 2025 preprint, "Veritas: Generalizable Deepfake Detection via Pattern-Aware Reasoning," combines the HydraFake benchmark with an end-to-end multimodal LLM trained to emit tagged reasoning patterns such as <fast>, <planning>, <reasoning>, <reflection>, and <conclusion> (Tan et al., 28 Aug 2025). HydraFake contains 50K real images and 50K fake images and evaluates in-domain, cross-model, cross-forgery, and cross-domain generalization (Tan et al., 28 Aug 2025). On this benchmark, Veritas reports 97.3% in-domain accuracy, 98.6% cross-model accuracy, 90.3% cross-forgery accuracy, 78.5% cross-domain accuracy, and 90.7% average accuracy; robustness experiments show 90.7 $\simeq 110\ \mathrm{m}^2$2 90.1 at JPEG QF=90, 88.7 at QF=70, 87.4 at QF=50, 88.8 at blur $\simeq 110\ \mathrm{m}^2$3, and 84.3 at $\simeq 110\ \mathrm{m}^2$4 (Tan et al., 28 Aug 2025).

The capitalization variant VeriTaS, expanded as Verified Theses and Statements, is a dynamic benchmark for multimodal automated fact-checking (Rothermel et al., 13 Jan 2026). It currently contains 24,000 real-world claims from 108 professional fact-checking organizations across 54+ languages, balanced into 1,000-claim quarterly splits from Q1 2020 to Q4 2025 (Rothermel et al., 13 Jan 2026). The benchmark is built by a fully automated seven-stage pipeline covering review discovery, publisher identification, article scraping, appearance retrieval, claim normalization, verdict standardization, and claim rectification (Rothermel et al., 13 Jan 2026). Verdicts are decomposed into four property scores—Media Authenticity, Media Contextualization, Claim Veracity, and Context Coverage—with overall Integrity defined as

$\simeq 110\ \mathrm{m}^2$5

Human evaluation on $\simeq 110\ \mathrm{m}^2$6 claims yields $\simeq 110\ \mathrm{m}^2$7, $\simeq 110\ \mathrm{m}^2$8, and 96.8% three-bin accuracy, and the benchmark is designed to remain leakage-resistant through quarterly updates (Rothermel et al., 13 Jan 2026).

Taken together, these later VERITAS systems are heterogeneous in application but unusually consistent in one respect: each formalizes evidence generation rather than treating the final output as sufficient. In the source papers, that evidence takes the form of fused expert scores, artifact-localized visual rationales, tagged reasoning traces, executable statistical outputs, or standardized, disentangled verdict properties.