🏛️ VeritasBoard — Multi-Agent Deliberative Quality Engine

Where artificial intelligence meets the rigor of a corporate boardroom, producing decisions that withstand the scrutiny of ten simulated experts and ten veteran reviewers.

🌟 Overview

VeritasBoard is an AI orchestration framework that reimagines quality control through simulated institutional wisdom. Rather than a single model generating output, VeritasBoard convenes two distinct chambers of artificial agents: a Drafting Council of ten specialist employees who produce preliminary work, and an Oversight Board of ten industry-veteran reviewers who challenge, refine, and validate that work. The user sees only the final, reconciled verdict — a distillation of twenty distinct perspectives, each engineered to resist sycophancy and groupthink.

This repository houses the core architecture, configuration templates, and integration modules that enable any language model to participate in structured, multi-agent deliberation. VeritasBoard is not merely a wrapper; it is a cognitive constitution — a set of rules that governs how artificial minds debate, dissent, and converge.

🧠 The Anti-Sycophancy Imperative

Modern language models suffer from a well-documented tendency: they agree with users. They seek approval. They hedge. VeritasBoard addresses this through deliberative adversarial review. Each of the ten Oversight Board members is prompted not to validate, but to interrogate — to find logical flaws, unsupported claims, hidden assumptions, and rhetorical gambits. The drafting agents, in turn, must defend their positions with evidence and structured reasoning.

The result is output that has been stress-tested by voices programmed to disagree.

🏛️ Architectural Pillars

Chamber 1: The Drafting Council (Ten Specialist Employees)

Chamber 2: The Oversight Board (Ten Industry Veterans)

🔄 Deliberation Workflow

The system operates in five stages, each with increasing levels of adversarial scrutiny:

Stage 1: Prompt Deconstruction (🕵️)

The user's query is parsed into its constituent claims, assumptions, and requests. Ambiguities are flagged for clarification.

Stage 2: Silent Drafting (✍️)

The ten drafting agents each produce an independent response. They do not communicate with each other — ensuring maximum diversity of approach.

Stage 3: Internal Review (🔍)

Each agent reviews their own draft for errors. The Critic agent performs a meta-analysis of all ten drafts, identifying overlapping strengths and isolated weaknesses.

Stage 4: Board Deliberation (⚖️)

The ten Oversight Board members receive the combined drafts and the Critic's meta-analysis. Each board member issues a verdict with specific recommendations. The Dissenting Voice must find something wrong, even if the output is excellent — forcing the others to articulate why they disagree with the dissent.

Stage 5: Reconciliation (🎯)

A final algorithm (configurable: majority vote, weighted consensus, or chairperson override) produces the single verdict the user sees. All dissenting opinions are logged but not displayed unless requested.

🌐 Multilingual & Cross-Cultural Deliberation

VeritasBoard operates in 17 languages as of January 2026, with culturally adapted prompting for each market. The Oversight Board's "Industry Veteran" personas include region-specific archetypes (e.g., a Japanese shacho for East Asian markets, a German Geschäftsführer for European contexts). This ensures that the anti-sycophancy mechanism does not inadvertently impose Western consensus norms on non-Western deliberation.

🎨 Responsive Interaction Modes

Each mode adjusts the temperature, presence penalty, and frequency penalty of underlying models to calibrate the balance between creativity and rigor.

🛠️ Configuration & Customization

Persona Builder

Define your own agent roles using a YAML schema:

agents:
  drafting_council:
    - name: "Domain Analyst"
      expertise: ["specialized_field"]
      instruction_tone: "critical_but_constructive"
  oversight_board:
    - name: "Chief Skeptic"
      default_stance: "challenge_every_claim"
      bias_correction: 0.95

Weighted Consensus Algorithm

Adjust how the final verdict is computed:

consensus:
  method: "weighted_majority"
  weights:
    chairperson: 2.0
    dissenting_voice: 1.5
    all_others: 1.0
  min_quorum: 12

📦 Module Index

📄 License Information

This project is released under the MIT License, granting permission to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the software, provided that the original copyright notice and permission notice appear in all copies.

View the full license text

⚠️ Disclaimer

VeritasBoard is a simulation framework designed for research, quality assurance, and structured output generation. The "agents" are language models with no independent cognition, intent, or legal standing. The outputs of VeritasBoard should not be considered as professional advice from actual human experts, nor should they replace human judgment in high-stakes environments such as medical diagnosis, legal representation, financial planning, or safety-critical decision-making.

As of 2026, VeritasBoard continues to evolve. While the anti-sycophancy mechanism drastically reduces model agreement bias, it cannot eliminate all forms of hallucination, subtle bias, or logical error. Always review critical outputs with human oversight.

🤝 Contributing & Community Ethos

We welcome contributions that strengthen the deliberative architecture. Areas of particular interest:

• Novel adversarial prompting techniques for the Dissenting Voice agent
• Cultural localization of board personas for non-Western markets
• Efficiency improvements that reduce token usage without sacrificing depth
• Audit logging for transparency in multi-agent decision paths

This project maintains a code of conduct that prioritizes constructive dissent — mirroring the very philosophy the framework embodies.