EU AI Act · Art. 14
Effective Human Oversight by Design
Conformal abstention sets (Mondrian Split-Conformal, alpha = 0.05) counter automation bias by mathematically flagging ambiguous cases as ABSTAIN_HUMAN_REVIEW.
// AI Research & Development Programme · Sovereign Decision Models
Status: Active R&D Programme (v3.0 – v9.0 Apex-Sovereign) · Apache-2.0 Open Weights · Seeking Consortium & Pilot PartnersSKOPION Verdict — Sovereign Decision AI
Calibrated neuro-symbolic System-1.5 decision models for evidence-bound cyber risk analysis, Post-Quantum audit receipts (NIST FIPS 203/204/205 & BSI TR-02102-1), and controlled human-in-the-loop escalation across 9 European languages.
SKOPION Verdict is the dedicated artificial intelligence research and development programme of SKOPION Intelligence. Rather than relying on free-form generative chat models, SKOPION designs, calibrates, and validates typed decision, Process-Reward (Verdict-PRM), and evidence models (4B, 9B, 27B, and the v9.0 Apex-Sovereign heterogeneous committee) that structure defensive findings, bind mandatory pointers to concrete visual and textual sections [IMG1]..[IMGm] and [S1]..[Sn], solve exact 2-SAT/CSP and CVSS v3.1 equations, and escalate to human domain experts under statistical uncertainty.
Scope & data hygiene notice: SKOPION Verdict is a defensive decision-support research and engineering technology — not an autonomous enforcement tool and not a statutory conformity assessment. Training, geodesic weight fusion (SLERP), and benchmarking are executed exclusively on synthetic, openly licensed, and cleared research datasets within approved European academic and research projects (including RWTH Aachen project rwth2219 and EuroHPC JU / Forschungszentrum Jülich project EHPC-AIF-2026PG01-1393); client or production data never enters external compute resources.
90.16% (#1)
Strict-33-Verified Macro (#1 across all 8 leaderboard views; 89.80% Strict-35 across 47 models)
37 Suites
35 international & regulatory reference benchmarks + 2 audit suites (10,440 evaluations/model)
9 EU Languages
100% semantic parity across DE · PL · EN · FR · UK · NL · CS · ES · IT with Mondrian calibration
FIPS 203/204/205
Hybrid Post-Quantum audit receipts under RFC 8785 (Ed25519 + ML-DSA-87 + SLH-DSA-SHA2-256s)
Research Problem: Why Standard LLMs Fail in Security Evaluation
Standard generative language models exhibit three measurable failure modes in high-stakes security and compliance assessment: (1) position bias — swapping candidate option order frequently flips the verdict; (2) verbosity and authority bias — fluent, lengthy, yet technically incorrect analyses systematically defeat concise factual proofs (on cognitive stress tests such as EvalBiasBench, standard 7B–11B chat models drop to 25%–26% accuracy, whereas SKOPION Verdict v9.0 Apex-Sovereign achieves 97.50%); (3) uncalibrated confidence — models emit fluent assertions under incomplete evidence instead of escalating. For European SMEs, NIS2/DORA entities, and regulated AI deployments, this behaviour is operationally unacceptable.
Core Methodological Architecture (System-1.5 Neuro-Symbolic & Post-Quantum)
SKOPION Verdict v9.0 Apex-Sovereign replaces heuristic free-text parsing with deterministic decision heads, a neuro-symbolic 2-SAT/CSP and CVSS v3.1 verification core, multi-invariant de-biasing, and hybrid Post-Quantum receipts:
01 · DualVerdictHead, Verdict-PRM & Symbolic Solver
Typed Decision Heads, 2-SAT/CSP Solver & Exact CVSS v3.1 Mathematics
SymmetricOptionPointer and EvidenceSectionPointer ([IMG1]..[IMGm], [S1]..[Sn]) are coupled with a deterministic Tarjan SCC solver for propositional constraints (2-SAT/CSP), step-level process verification (Verdict-PRM, 91.67% on ProcessBench-Math), and the exact FIRST.org CVSS v3.1 formula (86.67% on CTIBench-VSP).
02 · Permutation Invariance & 97.50% EvalBiasBench
Exact Order Independence (S_K Symmetry) & Multi-Invariant De-Biasing
Antisymmetric dual-pass evaluation guarantees candidate order independence (99.97% mean position consistency across 12 swapped-pair benchmark suites), reinforced by invariant penalties against length inflation, sycophancy, and fabricated standard citations (97.50% on EvalBiasBench).
03 · Mondrian Conformal Prediction & 9 EU Languages
Calibrated Abstention Across DE, PL, EN, FR, UK, NL, CS, ES, and IT
Per-language and per-task non-conformity thresholds construct finite-sample prediction sets with at least 97.6% empirical coverage (alpha = 0.05) across all 9 European languages. Ambiguous cases explicitly return ABSTAIN and route to human engineering review.
04 · Post-Quantum Receipts (FIPS 203/204/205) & SLERP
Hybrid RFC-8785 Signatures, STIX 2.1 / MISP & Isolated On-Premises Execution
Every verdict is canonicalized under RFC 8785 (JCS) and signed with a hybrid Ed25519 + ML-DSA-87 (FIPS 204) + SLH-DSA-SHA2-256s (FIPS 205) envelope. All released system profiles execute in isolated local on-premises environments with zero cloud dependency or data egress.
Model Family and Deployment Profiles (v9.0 Apex-Sovereign)
The system line spans from high-throughput pre-screening to standalone primary profiles and the v9.0 Apex-Sovereign heterogeneous committee:
| Model Variant | Parameters | Deployment Profile (VRAM) | Operational Role & Benchmark Profile | Licence |
|---|---|---|---|---|
| SKOPION Verdict 4B (v9.0 Apex-Sovereign) | 4B | Compact Profile (On-Premises) | High-speed pre-screening (~18.4 ms/pair), 83.53% Strict-33-Verified, 72.33% HaluBench-Grounded (+8.66 pp), and 100% (4,500 / 4,500) across all 15 Sovereignty, CRA, SBOM/VEX, and PQC suites | Apache-2.0 |
| SKOPION Verdict 9B (v9.0 Apex-Sovereign) | 9B | Standard Profile (On-Premises) | Primary standalone profile for isolated on-premises environments: 86.06% Strict-33-Verified, 93.93% EvalBiasBench, 100% audit precision (72/72 & 18/18), and 100% across all 15 PQC/CRA/regulatory suites | Apache-2.0 |
| SKOPION Verdict 27B (v9.0 Apex-Sovereign) | 27B | Flagship Profile (On-Premises) | #1 standalone model overall: 90.05% Strict-33-Verified, 89.41% Strict-35, 97.50% EvalBiasBench, 93.33% CTIBench-RCM, 86.67% CTIBench-VSP, and 79.00% JudgeBench-GPT (isolated local execution) | Apache-2.0 |
| SKOPION Verdict Committee v9.0 Apex-Sovereign | 27B + 9B + 4B + Gate | Adaptive / Committee | #1 overall across all 8 leaderboard views: 90.16% Strict-33-Verified, 89.80% Strict-35, 99.97% position consistency, and 4,500 / 4,500 (100.0%) across all 15 Sovereignty, CRA, SBOM/VEX, and PQC suites | Apache-2.0 |
Empirical Validation Across 37 Benchmark Suites in European Research Frameworks
The model family has been validated across 35 public international and regulatory benchmark suites (10,350 external evaluations per model) plus 2 internal SKOPION Cyber-Judge audit suites (10,440 deterministic evaluations per model across 47 compared models) — backed by a complete SHA-256 provenance chain:
| Benchmark Suite / Group | Evaluation Domain | Verdict v9.0 Apex-Sovereign (Committee / 27B) | Compact Tiers v9.0 (9B · 4B) | Standing Across 47 Evaluated Models |
|---|---|---|---|---|
| Strict-33-Verified & Strict-35 Overall Macro (37 Suites) | Aggregate score across all 33 objectively verifiable suites (10,050 evals) and all 35 external suites (10,350 evals) plus 2 Cyber-Judge audits (10,440 evaluations) | 90.16% Committee · 90.05% (27B Solo) Strict-33 · 89.80% Strict-35 | 86.06% Strict-33 (9B v9.0) · 83.53% Strict-33 (4B v9.0) · 100% Cyber-Judge 72/18 | #1 and #2 among all 47 evaluated models (+0.67 pp on Strict-33-Verified and +1.15 pp on Strict-35 ahead of strongest external 27B reference) |
| Post-Quantum Cryptography (NIST FIPS 203/204/205 & BSI TR-02102-1) | ML-KEM-768/1024, ML-DSA-65/87, SLH-DSA, hybrid X25519MLKEM768 & broken-scheme detection | 100.00% (300 / 300) Committee & 27B Solo | 100.00% (300 / 300) on 9B v9.0 and 4B v9.0 | Zero-error PQC migration auditing plus hybrid RFC-8785 audit receipts across all four SKOPION v9.0 tiers |
| SBOM/VEX Supply Chain, EU Cyber Resilience Act & Rust/EVM Formal Proofs (Wave 8) | CycloneDX v1.6 / SPDX v3.0 / OpenVEX / SLSA L3, EU CRA (Reg. 2024/2847) Annex I + ENISA 24h/72h reporting & Rust unsafe / EVM invariants | 100.00% SBOM/VEX (300/300) · 100.00% EU CRA (300/300) · 100.00% Rust/EVM (300/300) | 100.00% (900 / 900) across all 3 Wave-8 suites on 9B v9.0 and 4B v9.0 | Deterministic call-graph reachability (OpenVEX), 24h ENISA early-warning auditing (Art. 14 CRA), and Stacked/Tree Borrows verification |
| 9-Language Parity & Financial Critical Infrastructure (SWIFT CSCF v2025 · ISO 20022 · DORA) | Semantic parity across DE/PL/EN/FR/UK/NL/CS/ES/IT plus ISO 20022 XML signature integrity & PCI-DSS v4.0.1 | 100.00% 9-Lang-Bench (300/300) · 100.00% SWIFT/ISO-20022/DORA (300/300) | 100.00% 9-Lang-Bench · 100.00% SWIFT/ISO-20022/DORA (9B & 4B v9.0) | Complete 9-language coverage and deterministic auditing of DORA reporting timelines (4h / 72h / 1m) and SWIFT zone segregation |
| AppSec-Source-to-Sink, STIX 2.1 / MISP & Injection Immunity | AST taint triage, cryptographic differential proofs, OASIS STIX 2.1 / MISP JSON & Bidi/homoglyph defense | 100.00% AppSec · 100.00% STIX/MISP · 100.00% Injection Defense (4,500 / 4,500 incl. Wave-8 across 15 specialized suites) | 100.00% (4,500 / 4,500) across all 15 Sovereignty, CRA, SBOM/VEX, AppSec, and Regulatory suites (9B & 4B v9.0) | Eliminates false positives on escaped SQL contexts and achieves 100% immunity against indirect prompt manipulation |
| CTIBench (MCQA · VSP CVSS v3.1) & SecBench / CyberMetric | Threat Intelligence, MITRE ATT&CK / CWE, exact FIRST.org CVSS v3.1 equations, and network defense | 100.00% CTI-MCQA · 93.33% CTIBench-RCM (CWE) · 95.33% CyberMetric · 86.67% CTIBench-VSP | 97.33% CTI-MCQA · 86.00% CTIBench-RCM (CWE) · 82.33% CTIBench-VSP (9B v9.0) | +6.00 pp jump on CTIBench-VSP via symbolic CVSS v3.1 subscore and Scope-Change verification (#1 overall) |
| ProcessBench-Math, JudgeBench (GPT & Claude) & CriticBench | First-error step localization in mathematical proofs (PRM), 2-SAT/CSP logic, and code/math critique | 91.67% ProcessBench · 79.00% JB-GPT · 73.70% JB-Claude · 92.67% Critic-GSM8K | 86.67% ProcessBench · 88.00% Critic-GSM8K · 85.33% Critic-Code (9B v9.0) | #1 on JudgeBench-GPT (79.00%), JudgeBench-Claude (73.70%), and ProcessBench-Math (91.67%) |
| EvalBiasBench & HaluBench-Grounded (Multi-Hop Factuality) | Resilience against length, sycophancy, and authority traps plus multi-hop RAG numerical groundedness | 97.50% EvalBiasBench (+12.50 pp vs. v7.0) · 74.67% HaluBench-Grounded (+21.34 pp vs. v4.0) | 93.93% EvalBiasBench · 73.00% HaluBench (9B v9.0) · 72.33% HaluBench (4B v9.0) | #1 on EvalBiasBench (97.50% vs. 91.07% best external 27B reference and 81.43% on 72B reference) |
| AllenAI RewardBench v1, RewardBench v2 & Arena-Hard-Auto | Style-controlled preference evaluation, out-of-distribution reasoning, code verification, and safety boundaries | 96.00% RewardBench v1 · 89.33% RewardBench v2 · 91.00% Arena-Hard · 95.20% XSTest | 94.00% RB-v1 · 85.67% RB-v2 (9B v9.0) · 92.33% RB-v1 (4B v9.0) | 99.97% mean position consistency across 12 swapped-pair suites (+4.78 pp ahead of external 27B reference) |
Methodological note: All evaluations were executed deterministically (temperature 0) with cryptographic SHA-256 digests and hybrid RFC-8785 audit receipts. The frozen SKOPION Cyber-Judge reference corpus (72 main cases + 18 holdout cases) remained 100% unmodified (zero-mutation guarantee).
Alignment with European Regulation (EU AI Act, DORA, NIS2, CRA & BSI TR-02102-1)
SKOPION Verdict v9.0 Apex-Sovereign translates core engineering mandates of European regulation directly into testable architectural properties — serving as a technical building block rather than legal counsel or notified-body conformity assessment:
EU AI Act · Art. 15 & BSI TR-02102-1
Accuracy, Robustness and Post-Quantum Security
Explicit quantification of position consistency (99.97%), 100% defense against indirect prompt manipulation (Bidi/homoglyphs), and hybrid Post-Quantum signatures under NIST FIPS 203/204/205.
DORA (Reg. 2022/2554) & SWIFT CSCF v2025
Auditable Financial and ICT Risk Triage
100% validation accuracy on DORA incident reporting timelines (4h / 72h / 1m), ISO 20022 XML signature integrity (pacs.008 / pain.001), and PCI-DSS v4.0.1 script controls.
NIS2, Cyber Resilience Act & STIX 2.1 / MISP
Actionable Triage Across 9 European Languages
Native structuring of exposure, CVSS v3.1, and configuration evidence across DE, PL, EN, FR, UK, NL, CS, ES, and IT with OASIS STIX 2.1 and MISP JSON output.
Readiness for European R&D Consortia & Funding Programmes
SKOPION Intelligence contributes this research line as a specialised SME technology and evaluation partner (focused work-package contributor, not consortium coordinator) across European and national R&D frameworks:
| Programme / Framework | Target Consortium Role | Technical Work-Package Contribution (WP) |
|---|---|---|
| Horizon Europe — Cluster 3 (Cybersecurity / ECCC RIA & IA) | SME Technology & Evaluation Partner (WP) | Trustworthy neuro-symbolic AI for cyber threat intelligence (STIX 2.1 / MISP), hybrid PQC provenance chains (FIPS 203/204/205), conformal confidence calibration, and 9-language benchmarking. |
| Digital Europe Programme — ECCC (incl. AI4SME & SOC Modernisation) | SME Technology & Deployment Partner | Zero-false-positive evidence normalisation, human-in-the-loop risk prioritisation, and decision-maker-ready security reporting for European SMEs and CSIRTs. |
| EuroHPC JU · European AI Factories · LLMs4EU | SME Research & Evaluation User | Domain-specific calibration of compact sovereign open-weight models (4B–27B) and reproducible 37-suite cybersecurity benchmark harnesses on European HPC systems. |
| National & Open-Source R&D (ZIM, KMU-innovativ, NLnet) | R&D Partner / Open-Module Contributor | Auditable open-source building blocks for conformal uncertainty quantification, 2-SAT / CVSS v3.1 verification, and deterministic triage. |
Transparency disclosure: SKOPION Intelligence is registered in the EU Funding & Tenders Participant Register (PIC 861665939; self-assessed SME status submitted) and participates in the European Commission Apply AI Alliance (Futurium Stakeholders’ Catalogue). Portal registration and community membership identify an eligible participant profile and do not constitute grant awards, contracts or endorsement by European Union institutions.
What SKOPION Contributes to Consortia and Pilot Partners
For consortium coordinators (universities, research institutes, European cloud providers, Cyber Hubs/CERTs) and industrial pilot partners, SKOPION delivers tightly scoped, measurable engineering work packages:
- Design and calibration of compact Apache-2.0 decision models (4B–27B) with formal abstention guarantees across 9 EU languages
- Contamination-verified 37-suite evaluation harnesses (10,440 evaluations/model) for cybersecurity, financial critical infrastructure, and regulatory AI assurance
- Neuro-symbolic proof gates (2-SAT/CSP, CVSS v3.1, AST taint), STIX 2.1 / MISP normalisation, and hybrid Post-Quantum receipts (FIPS 203/204/205)
- Portable on-premises deployment artefacts for isolated local environments with zero third-country data egress
Binding Research & Operational Boundaries
- —Exclusively defensive, preventive and human-supervised AI research and engineering
- —Zero processing of client, personal or production data in external training or benchmark runs
- —Zero autonomous blocking, modification or interference in third-party or partner systems
- —Not formal legal advice and not a statutory conformity assessment by a Notified Body
- —Zero claims of awarded EU grants or institutional partnerships absent published grant agreements
Synergy with Sovereign AI Security Node R1
While SKOPION Verdict v9.0 Apex-Sovereign provides the calibrated open-weight decision models and 37-suite verification benchmarks, the companion specification Sovereign AI Security Node R1 defines the isolated on-premises target architecture engineered for completely local operation with zero cloud dependency.