Model Transparency Report
Comprehensive documentation of the AI systems powering Aevum Encyclopedia's knowledge retrieval, synthesis, and verification pipelines.
1. Executive Summary
This report outlines the architecture, training methodology, evaluation metrics, safety protocols, and governance framework for Aevum-Knowledge-3, the primary large language model (LLM) powering Aevum Encyclopedia's AI-assisted search, content synthesis, citation verification, and multilingual translation features.
Aevum Encyclopedia is committed to radical transparency regarding our AI systems. This document is intended for researchers, educators, compliance auditors, and the general public. All claims are supported by internal evaluations and third-party audits where applicable.
This report covers the base knowledge model and its production fine-tunes. It does not cover third-party integrations or legacy rule-based systems still in maintenance mode.
2. Model Architecture & Capabilities
Aevum-Knowledge-3.2 is a sparse mixture-of-experts (MoE) transformer optimized for factual retrieval, long-context reasoning, and academic-grade synthesis.
| Parameter | Specification |
|---|---|
| Architecture | Dense + MoE Hybrid (32 Experts, 8 Activated) |
| Total Parameters | 137B (Active: 14.2B) |
| Context Window | 128K tokens |
| Vocabulary | Custom BPE + WordPiece (256K tokens) |
| Inference Framework | vLLM + Custom KV-Cache Optimizer |
| Latency (P95) | 320ms (first token), 850ms (full response) |
| Supported Modalities | Text, Structured Data, Metadata Graphs |
Core Capabilities
- Cross-Document Synthesis: Aggregates verified claims across thousands of sources while preserving attribution.
- Fact-Verification Pipeline: Real-time cross-referencing against a curated knowledge graph and primary source archive.
- Multilingual Grounding: Native training across 140+ languages with culture-aware terminology mapping.
- Academic Formatting: Automatic generation of APA, MLA, Chicago, and IEEE citation structures.
3. Training Data & Provenance
Training data was curated using a multi-stage filtering pipeline designed to prioritize academic rigor, copyright compliance, and demographic balance.
Data Sources & Composition
| Category | Proportion | Examples | Licensing |
|---|---|---|---|
| Open Academic & Preprints | 42% | arXiv, PubMed, OpenAlex, DOAJ | CC-BY, Open Access |
| Encyclopedia & Reference | 28% | Wikipedia, Britannica (licensed), Stanford Encyclopedia | CC-BY-SA, Commercial |
| Gov & Institutional Reports | 18% | UN, WHO, NOAA, OECD, National Archives | Public Domain |
| Curated Media & Journalism | 8% | Reuters, AP, Associated Press Archive | Licensed |
| Synthetic & Augmented | 4% | Controlled QA pairs, citation exercises | Proprietary |
Preprocessing & Filtering
- Deduplication: MinHash + LSH fingerprinting reduced exact/near-duplicate content by 34%.
- Toxicity & Bias Screening: Multi-layer classifier removed content exceeding safety thresholds (F1: 0.94 on detoxified benchmarks).
- PII Removal: NER-based redaction pipeline achieved 99.2% precision on personal identifiers.
- Licensing Verification: Automated crawler + legal review board validated usage rights for all commercial datasets.
4. Evaluation & Benchmarks
The model undergoes continuous evaluation across factual accuracy, hallucination rates, bias metrics, and safety alignment.
| Benchmark | Score | Industry Avg | Status |
|---|---|---|---|
| MMLU (Academic) | 89.4% | 82.1% | Above |
| TruthfulQA | 91.2% | 78.5% | Above |
| FACTS-Citation Recall | 96.8% | 88.0% | Above |
| Hallucination Rate (Open QA) | 2.1% | 6.4% | Below |
| Cross-Cultural Bias (CC-Bias) | 0.18 | 0.34 | Monitored |
Independent evaluation by AI Governance Institute (AIGI) completed Q2 2025. Full technical annex available upon request for accredited researchers.
5. Risk Assessment & Mitigations
We proactively identify and document systemic risks associated with model deployment in an educational and research context.
Identified Risks
| Risk Category | Severity | Mitigation Strategy |
|---|---|---|
| Factual Hallucination | High | Real-time citation gating; refusal to answer unverified claims; confidence scoring |
| Cultural/Linguistic Bias | Medium | Curated multilingual fine-tuning; periodic bias audits; regional expert review panels |
| Academic Misuse | Medium | Plagiarism-aware output markers; institutional API watermarking; educational guidelines |
| Adversarial Prompting | Low | Input sanitization layer; rate limiting; behavior anomaly detection |
Safety Filters
- Pre-Processing: Prompt classification blocks harmful, deceptive, or overly broad queries.
- In-Context: Confidence thresholds trigger fallback to reference sources or clarification requests.
- Post-Processing: Output scanning flags unverified claims, missing citations, or tone inconsistencies before delivery.
6. Governance & Oversight
Aevum Encyclopedia maintains a multi-tiered governance structure to ensure responsible AI development and deployment.
Organizational Structure
- AI Ethics Board: Independent panel of academics, legal experts, and civil society representatives.
- Red-Team Unit: Internal group conducting monthly adversarial testing and edge-case discovery.
- Community Review Panel: Volunteer subject-matter experts who flag discrepancies and suggest updates.
Compliance Standards
Our systems align with the EU AI Act (Category II: Limited Risk), NIST AI Risk Management Framework, and UNESCO Recommendation on the Ethics of AI. All model cards, data sheets, and impact assessments are version-controlled and publicly accessible.
7. Version History & Updates
| Version | Date | Key Changes | Status |
|---|---|---|---|
| 3.2 | 2025-03-10 | Enhanced citation verification; reduced hallucination rate by 14%; expanded Arabic/Indonesian support | Current |
| 3.1 | 2024-11-20 | MoE architecture rollout; context window extended to 128K; academic formatting module | Legacy |
| 3.0 | 2024-06-15 | Base model release; 140-language support; initial safety filters | Deprecated |
Major updates are preceded by 30-day public notice and compatibility testing. Deprecation cycles follow a 180-day sunset policy for enterprise APIs.
8. Contact & Feedback
We encourage researchers, educators, and users to submit feedback, report discrepancies, or request additional transparency documentation.
| Channel | Purpose | Response Time |
|---|---|---|
| transparency@aevumencyclopedia.com | General inquiries, documentation requests | 3-5 business days |
| responsible-ai@aevumencyclopedia.com | Safety concerns, bias reports, ethical feedback | 48 hours |
| Research Portal (aevum.dev/research) | Benchmark datasets, model cards, technical annexes | Self-service |
All verified reports are logged in our internal incident tracker. Publicly actionable findings are documented in monthly transparency digest updates.