VERITAS OS: Auditable Decision Operating System for Enterprise-Grade LLM Agents

VERITAS OS is a decision operating system created by Japanese developer Takeshi Fujishita, designed specifically for LLM Agents. Its core focus is governance (auditable, secure, compliant). It enables AI decision-making to be reviewable, traceable, and accountable through a pipeline of over 20 stages, FUJI Security Gates, hash-chain TrustLog, etc. It is suitable for regulated high-risk fields such as finance and healthcare. The current version is 2.0.0 Beta, with the positioning: "LLM is the CPU, VERITAS OS is the decision/Agent operating system."

VERITAS OS differs from Agent frameworks that pursue autonomy; its optimization direction is governance: auditable decision pipeline, fail-closed security mechanism, enterprise-level compliance risk control. Applicable scenarios include regulated industries (finance, healthcare, law), high-risk decision-making scenarios, multi-Agent collaboration systems, enterprise compliance needs, and AI security research.

Each decision must go through a structured pipeline process, with steps including input standardization, memory retrieval, web search, option standardization, core execution, result absorption, alternative generation, model enhancement, debate, critique, FUJI pre-check, value core evaluation, gated decision-making, value learning, metric calculation, evidence reinforcement, response assembly, persistence (audit + memory + world model), evidence finalization, and building a replay snapshot. This ensures decisions have complete context and audit trails.

FUJI Gate: Fail-closed design (returns rejection on anomalies). Multi-layer protection includes PII detection, harmful content interception, sensitive field handling, prompt injection defense, etc. On anomalies, it returns a rejected status and a risk=1.0 score. TrustLog: Hash-chain tamper-proof audit log. Features include RFC8785 JSON, SHA256 chain hash, Ed25519 signature, WORM hard-fail mirroring, etc. Data protection includes PII desensitization and AES-256-GCM encryption.

ValueCore: 14-dimensional value function (9 core ethics +5 policy-level), which learns from TrustLog through mechanisms like EMA weight update and automatic rebalancing. MemoryOS: Unified storage for four types of memory: situational, semantic, procedural, and emotional. Supports vector search and PII masking. WorldModel: Maintains the Agent's perception of the world state (snapshots, causal tracking, etc.). Debate and critique mechanism: multi-perspective reasoning and self-review to reduce bias.

Mission Control: Next.js 16 front-end dashboard, including homepage (real-time event stream), decision console, audit browser, governance control, risk dashboard, etc. Security design uses BFF mode, httpOnly Cookie, etc. Compliance: EU AI Act ready (compliance reports, audit exports). Enterprise governance includes four-eye approval, RBAC/ABAC, SSE alerts, etc. The Replay engine supports high-fidelity decision replay (non-strictly deterministic).

Backend: Python3.11+, FastAPI. Supports multiple LLM providers (default OpenAI GPT-4.1-mini). Frontend: Next.js16, React18, TypeScript5.7, etc. Deployment: Docker Compose (recommended), backend-only Docker, local development (make dev-all).

VERITAS passed an independent technical due diligence on 2026-03-15, receiving an A- rating (overall 82/100). License tiers: core is proprietary EULA, interfaces are MIT. Conclusion: VERITAS represents the "governance-first" direction of Agent infrastructure. Although it is a Beta version, its architecture is complete. As AI regulations are implemented, it will become the foundation for responsible deployment.